Roadmap to Neuromorphic Computing with Emerging Technologies

1. Introduction to the roadmap

Adnan Mehonic, Daniele Ielmini, Kaushik Roy

1.1. Taxonomy & Motivation

The growing adoption of data-driven applications, such as artificial intelligence (AI), is transforming the way we interact with technology. Currently, the deployment of AI and machine learning tools in previously uncharted domains generate considerable enthusiasm for further research, development, and utilisation. These innovative applications often provide effective solutions to complex, longstanding challenges that have remained unresolved for years. By expanding the reach of AI and machine learning, we unlock new possibilities and facilitate advancements in various sectors. These include but are not limited to, scientific research, education, transportation, smart city planning, eHealth, and the metaverse.

However, our predominant focus on performance can sometimes lead to critical oversights. For instance, our constant dependence on immediate access to information might cause us to ignore the energy consumption and environmental consequences associated with the computing systems that enable such access. Balancing performance with sustainability is crucial for the technology’s continued growth.

From this standpoint, the environmental impact of AI is a cause for growing concern. Additionally, applications such as the Internet of Things (IoT) and autonomous robotic agents may not always rely on resource-intensive deep learning algorithms but still need to minimize energy consumption. Realizing the vision of IoT is contingent upon reducing the energy requirements of numerous connected devices. Demand for computing power is growing at a rate that far exceeds improvements achieved through Moore’s law scaling. Figure 1a shows the computing power demands, quantified in Peta floating-point operations (PetaFLOPS, one peta = 10¹⁵) per day, as a function of time, indicating an increase of a factor 2 every two months in recent years ¹. In addition to Moore’s law, significant advancements have been made through the combination of intelligent architecture and hardware-software co-design. For instance, NVIDIA GPUs’ performance has improved by a factor of 317 from 2012 to 2021, surpassing expectations based on Moore’s law alone. Research and development efforts have demonstrated further impressive performance improvements ², ³, ⁴ suggesting that more can be achieved. However, conventional computing solutions alone are unlikely to meet demand in the long term, particularly when considering the high costs of training associated with the most complex deep learning models (Figure 1b). It is essential to explore alternative approaches to tackle these challenges and ensure the long-term sustainability of AI’s rapid advancements. While global energy consumption is crucial and important, there is a relevant issue which is perhaps just as significant: the ability of low-power systems to execute complex AI algorithms without relying on cloud-based computing. It is important to keep in mind that the challenge of global AI power consumption and the ability to implement complex AI on low-power systems are two somewhat separate challenges. It might be the case that these two challenges need to be addressed with somewhat different strategies (e.g., the power consumption in data centers for the most complex, largest AI models, such as large language models, might be addressed differently than implementing mid-sized AI models, such as voice recognition, on low-power, self-contained systems that might need to run at a few milliwatts of power). The latter strategy might not be scalable for the largest models, or the optimization of the largest models might not be applicable for simpler models running on much lower power budgets. However, undeniably, for both, we need to improve the overall energy efficiency of our computing systems that are designed to execute AI workloads.

Figure 1: a) Increase in computing power demands to run state-of-the-art AI models. b) The cost associated with training AI models.  Adapted and reproduced from ¹ .

The energy efficiency and performance of computing can largely benefit from new paradigms that aim at replicating or being inspired by specific characteristics of the brain’s biological mechanisms. It is important to note that biological systems might be highly specialized and heterogenious, and therefore different tasks are addressed by different computational schemes. However, we can still aim to take inspiration from general features when they are advantageous for specific applications. It is unlikely that a single architecture or broader approach will be best applicable for all targeted applications.

Adopting an interdisciplinary methodology, experts in materials science, device and circuit engineering, system design, and algorithm and software development are brought together to collectively contribute to the progressive field of neuromorphic engineering and computing. This collaborative approach is instrumental in fueling innovation and promoting advancements in a domain that seeks to bridge the gap between biological systems and artificial intelligence. Coined by Carver Mead in the late 1980s ⁵, the term ‘neuromorphic’  originally referred to systems and devices replicating certain aspects of biological neural systems, but now it varies across different research communities. While the term’s meaning continues to evolve, it generally refers to a system embodying brain-inspired properties such as in-memory computing, hardware learning, spike-based processing, fine-grained parallelism, and reduced precision computing, among others. One can also draw analogies and identify more complex phenomenological similarities between biological units (e.g., neurons) and electronic components (e.g., memristors). For example, phenomenological similarities between models of the redox-based nanoionic resistive memory cell and common neuronal models, such as Hodgkin-Huxley conductance model and the leaky integrate-and-fire model, have been demonstrated ⁶. Even more complex biological functionalities have been demonstrated using a single third-order nanocircuit elements ⁷. It should be noted that many paradigms related to the neuromorphic approach have also been independently investigated. For instance, in-memory computing ⁸, while being a cornerstone of the neuromorphic paradigm, is also examined separately. It represents one of the most promising avenues to enhance the energy efficiency of AI hardware or more general computing, offering a break from the traditional Von Neumann architecture paradigm.

Neuromorphic research can be divided into three areas. Firstly, “neuromorphic engineering” employs either CMOS technology (e.g. transistors working in a sub-threshold regime) or cutting-edge post-CMOS technologies to reproduce the brain’s computational units and mechanisms. Secondly, “neuromorphic computing” explores new data processing methods, frequently drawing inspiration from biological systems and considering alternative algorithms, such as spike-based computing. Lastly, the development of “neuromorphic devices” marks the third field. Taking advantage of advancements in electronic and photonic technologies, it develops innovative nano-devices that frequently emulate biological components like neurons and synapses or efficiently implement desired properties, such as in-memory computing.

Furthermore, various approaches to neuromorphic research can be identified based on their primary objectives. Some systems focus on delivering efficient hardware platforms to enhance our understanding of biological nervous systems, while others employ brain-inspired principles to create innovative, efficient computing applications. This roadmap primarily focuses on the latter. While there are already outstanding roadmaps ⁹, reviews ¹⁰,  ¹¹, ¹² and the special issues ¹³ that offer comprehensive overviews of neuromorphic technologies, encompassing the integration of hardware and software solutions as well as the exploration of new learning paradigms, this particular roadmap focuses on emphasizing the significance of materials engineering in advancing cutting-edge complementary metal-oxide-semiconductor (CMOS) and post-CMOS technologies. Simultaneously, it offers a holistic perspective on the general challenges of computing systems, the reasoning behind adopting the neuromorphic approach, and concise summaries of current technologies to better contextualize the role of materials engineering within the broader neuromorphic landscape. Of course, there are other critical aspects in the development of neuromorphic technologies that need to be taken into account. For example, an excellent recent review on thermal management materials, devices, and networks is one such example ¹⁴.

The roadmap is organized into several thematic sections, outlining current computing challenges, discussing the neuromorphic computing approach, analyzing mature and currently utilized technologies, providing an overview of emerging technologies, addressing material challenges, exploring novel computing concepts, and finally examining the maturity level of emerging technologies while determining the next essential steps for their advancement.

The roadmap starts with a concise introduction to the current digital computing landscape, primarily characterized by Moore’s law scaling and the Von Neumann architecture. It then explores the challenges in sustaining Moore’s law and examines the significance and potential advantages of post-CMOS technologies and architectures aiming to integrate computing and memory. Following this, the roadmap presents a historical perspective on the neuromorphic approach, emphasizing its potential benefits and applications. It provides a thorough review of cutting-edge developments in various emerging technologies, comparing them critically. The discussion addresses how these technologies can be utilized to develop computational building blocks for future computing systems. The roles of two mature technologies, static random access memory (SRAM) and Flash, are also explored. The overview of emerging technologies includes resistive switching and memristors, phase change materials, ferroelectric materials, magnetic materials, spintronic materials, optoelectronic and photonic materials, and 2D devices and systems. Material challenges are discussed in detail, covering types of challenges, possible solutions, and experimental techniques to study these. Novel computing concepts are examined, focusing on embracing device and system variability, spiking-based computing systems, analog computing for linear algebra, and the use of analog content addressable memory (CAM) for in-memory computing and optimization solvers. The final section discusses technological maturity and potential future directions.

2. Computing challenges

Onur Mutlu, Shahar Kvatinsky

2.1. Digital computing

2.1.1. Status 

Digital computing has a long and complex history that stretches back over a century. The earliest electronic computers were developed in the 1930s and 1940s, and they were large, expensive, and difficult to use. However, these early computers laid the foundation for the development of the modern computers that we use today and their principles are still in widespread use.

One of the key figures in the early history of digital computing was John von Neumann, a mathematician and computer scientist known for his contributions to the field of computer science. Von Neumann advocated the stored program concept and sequential instruction processing, two vital features of the von Neumann architecture ¹⁵ that are still used in most computers today. Another key feature of the von Neumann architecture is the separation of the CPU (control unit) and the main memory. This separation allows the CPU to access the instructions and data it needs from the main memory while executing a program, and assigns the computation and control responsibilities specifically to the CPU.

Throughout the years, the rapid scaling of semiconductor logic technology, known as Moore’s law ¹⁶,  has led to tremendous improvements in computer performance and energy efficiency. With the exponential increase in the number of transistors placed on a single chip provided by technology scaling, engineers have explored many ways to increase the speed and performance of computers. One way they did this was by exploiting parallelism, which is the ability of a computer to perform multiple tasks simultaneously. There are several different types of parallelism, including SISD (single instruction, single data), SIMD (single instruction, multiple data), MIMD (multiple instruction, multiple data), and MISD (multiple instruction, single data) ¹⁷, all of which are exploited in modern computing systems ranging from general-purpose single-core and multi-core processors, GPUs, and specialized accelerators.

Technology scaling has also allowed for the development of more processing units, starting from duplicating the processing cores and, more recently, adding accelerators. These accelerators can offload specific tasks, e.g., video processing, compression/decompression, vision processing, graphics, and machine learning, from the central processor, further improving performance and energy efficiency (the required energy to perform a certain task) by specializing the computation units to the task at hand. As such, modern systems are heterogeneous, with many different types of logic-based computation units integrated into the same processor die.

2.1.2. Challenges

While the performance and energy of logic-based computation units have scaled very well via technology scaling, those of interconnect and memory systems have not scaled as well. As a result, communication (e.g., data movement) between computation units and memory units has emerged as a major bottleneck, partly due to the disparity in scaling and partly due to the separation and disparity between processing and memory offered in von Neumann architecture, which both have limited the ability of computers to take full advantage of the improvements in logic technology. This bottleneck is broadly referred to as the “von Neumann bottleneck” or the “memory wall,” as it can greatly limit the speed and energy at which the computer can execute instructions.

For decades, the transistor size has scaled down while the power density has remained constant. This phenomenon, first observed in the 1970s by Robert Dennard ¹⁸, means that as transistors become smaller and more densely packed onto a chip, the overall performance and capabilities of the chip improve. However, since the early 2000s, it has become increasingly challenging to maintain Dennard scaling as voltage (and thus frequency) scaling has greatly slowed down. The end of Dennard scaling has increased the importance of energy efficiency of different processing units and led to phenomena such as “dark silicon,” ¹⁹ where large parts of the chip are powered off. The rapid move towards more specialized processing units, powered on for specific tasks, exemplifies the influence of the end of Dennard scaling.

Furthermore, in recent years, it has become increasingly challenging to maintain the pace of Moore’s law due to the physical limitations of transistors and the challenges of manufacturing smaller and more densely packed chips. As a result, the looming end of Moore’s law has been a topic of discussion in the tech industry, as this could potentially limit the future performance improvements of computer chips. New semiconductor technologies and novel architectural solutions are required to continue computing systems’ performance and energy efficiency improvements at a similar pace as in the past.

2.1.3. Potential Solutions and Conclusion

In recent years, different semiconductor and manufacturing technologies have emerged to overcome the slowdown of Moore’s law. These devices include new transistor structures and materials, advanced packaging techniques, and new (e.g., nonvolatile) memory devices. Some of those technologies have similar functionality as standard CMOS technology but with improved properties. Other technologies also offer radically new properties, different from CMOS. For example, memristive technologies, such as resistive RAM ²⁰, have varying resistance and provide analog data storage that also supports computation. Such novel technologies with their unique properties may serve as enablers for new architectures and computing paradigms, which could be different from and complementary to the von Neumann architecture.

The combination of Moore’s law slowdown and von Neumann’s bottleneck requires fresh thinking on computing paradigms. Data movement between the memory and the processing units is the primary impediment against high performance and high energy efficiency in modern computing systems ^21,22,23,24. And, this impediment only worsens with the improved processing abilities and the increased need for data. All modern computers employ a variety of methods to mitigate the memory bottleneck, all of which increase the complexity and power requirements of the system with limited (and sometimes little) success in mitigating the bottleneck. For example, modern computers have several levels of cache memories to reduce the latency and power of memory accesses by exploiting data locality. Cache memories, however, have limited capacity and are effective only when significant spatial and temporal locality exists in the program. Cache memories are not always (completely) effective due to low locality in many modern workloads, which can worsen the performance and energy efficiency of computers ^25,26. Similarly, modern computers employ prefetching techniques across the memory hierarchy to anticipate future memory accesses and load data into caches before it is needed by the processor. While partially effective for relatively simple memory access patterns, prefetching is not effective for complicated memory access patterns and it increases system complexity and memory bandwidth consumption ²⁷. Thus, memory bottleneck remains a tough challenge and hundreds of research papers and patents are written every year to mitigate it ²⁸.

Overcoming the performance and energy costs of off-chip memory accesses is an increasingly difficult task as the disparity between the efficiency of computation and efficiency of memory access continues to grow. There is therefore a need to examine more disruptive technologies and architectures that much more tightly integrate logic and memory at a large scale, avoiding the large costs of data movement across system components.

Many efforts to move computation closer to and inside the memory units have been made ²⁹, including adding processing units in the same package as DRAM chips ^30,31, performing digital processing using memory cells 10.1038/nature08940 ^32,33, and using analog computation capabilities of both DRAM and NVM devices ^34,35,36,37. One exciting novel computing paradigm to eliminate the von Neumann bottleneck is to reconsider the way computation, and memory tasks are performed by getting inspiration from the brain, where, unlike von Neumann architecture, processing and storage are not separated. Many recent works demonstrate orders of magnitude performance and energy improvements using various kinds of processing-in-memory architectures ²⁹. Processing-in-memory and, more broadly, neuromorphic (or brain-inspired) computing thus offers a promising way to overcome the major performance and energy bottleneck in modern memory systems. However, it also introduces significant challenges for adoption as it is a disruptive technology that affects all levels of the system stack, from hardware devices to software algorithms.

3. Neuromorphic computing basics and its evolution

Teresa Serrano-Gotarredona, Bernabe Linares Barranco

3.1 What is neuromorphic computing/engineering

Neuromorphic computing can be defined as the underlying computations performed by neuromorphic physical systems. Neuromorphic physical systems carry out robust and efficient neural computation using hardware implementations that operate in physical time. Typically, they are event- or data-driven, and typically they employ low-power, massively parallel hybrid analog, digital or mixed VLSI circuits, and they operate using similar physics of computation used by the nervous system .

Spiking neural networks (SNN) are one very good example of a neuromorphic computing system. Computation is performed whenever a spike is transmitted and received by destination neurons. Computation can be performed at the dendritic tree, while spikes travel to their destinations, as well as at the destination neurons where they are collected to update the internal states of the neurons. Neurons collect pre-weighted and pre-filtered spikes coming from different source neurons or sensors, perform some basic computation on them and generate an output spike whenever their internal state reaches some threshold. A neuron firing typically means the “feature” this neuron represents has been identified in place and time. The collective computation of populations of neurons can give rise to powerful system level behavior, such as pattern recognition, decision making, sensory fusion, knowledge abstraction, and so on. Additionally, neuromorphic computing systems can also be enabled to acquire new knowledge through both supervised and unsupervised learning, either off-line or while they perform, which is typically known as on-line learning and which can be life-long. Neuromorphic computing covers typically from sensing to processing to learning.

3.1.1. Neuromorphic Sensing

Probably the most clarifying example of what neuromorphic computation is about, is the paradigm of neuromorphic visual computation. Neuromorphic visual computation exploits the data encoding provided by neuromorphic visual sensors. Today, the most wide-spread neuromorphic vision sensor is the Dynamic Vision Sensor (DVS) [³⁸]. In a DVS each pixel sends out its (x,y) coordinate whenever its photodiode perceives a relative change of light beyond some pre-set thresholds $\theta^{-}>\frac{I_{n+1}}{I_{n}}>\theta^{{}_{+}}$θしーた start_POSTSUPERSCRIPT - end_POSTSUPERSCRIPT > divide start_ARG italic_I start_POSTSUBSCRIPT italic_n + 1 end_POSTSUBSCRIPT end_ARG start_ARG italic_I start_POSTSUBSCRIPT italic_n end_POSTSUBSCRIPT end_ARG > italic_θしーた start_POSTSUPERSCRIPT start_FLOATSUBSCRIPT + end_FLOATSUBSCRIPT end_POSTSUPERSCRIPT, with $\theta^{+}$θしーた start_POSTSUPERSCRIPT + end_POSTSUPERSCRIPTslightly greater than 1 and $\theta^{-}$θしーた start_POSTSUPERSCRIPT - end_POSTSUPERSCRIPTslightly less than 1. This is typically referred to as an “Address Event”. If  $I_{n+1}>I_{n}$then light has increased. If $I_{n+1}<\ I_{n}$then light has decreased. To differentiate both situations, the address event can also be a signed event, by adding a sign bit ‘s’, (x,y,s). If events are recorded using some event-recording hardware, then a timestamp t_n is added to each event (x_n,y_n,s_n,t_n). The full recording consists then of a list of time-stamped address events. Fig. 2 illustrates this. In Fig. 2(a) a DVS camera is observing a 7KHz spiral on a classic phosphor oscilloscope (without any extra illumination source). Fig. 2(b) plots in {x,y,t} space the recorded events. The camera was a 128x128 pixel high-contrast sensitivity DVS camera [³⁹]. Therefore, x-y coordinates in  Fig. 2(b) spawn from 0 to 127. The vertical axis is time, which spawns over about 400us, slightly less than 100us per spiral turn. Each dot in Fig. 2(b) is an address event, and we can count several hundreds of them within the 400us. This DVS camera is capable of generating over 10 million events per second (about one every 100ns). This produces a very fine timing resolution when sensing dynamic scenes.

The information (events) produced by this type of sensors can be sent directly to event-driven neuromorphic computing hardware, which would process this quasi-instantaneous dynamic visual information event by event.

Figure 2: (a) 7KHz spiral observed in a classic phosphor oscilloscope set in X/Y mode. (b) DVS output event stream when observing the oscilloscope in (a).

DVS cameras have evolved over the past 20 years, since they first appeared [⁴⁰], combined with frames, sensitive to color [⁴¹], and of resolutions up to 1-Mega-pixel [⁴²].

Other sensory modality event-driven neuromorphic devices have been reported, such as auditory cochleae [⁴³], olfactory noses [⁴⁴], or tactile sensing [⁴⁵].

3.1.2. Neuromorphic Processing

Neuromorphic signal information encoding in the form of sequences of events reduces information so that only meaningful data, such as changes, are transmitted and processed. This follows the underlying principle in biological nervous systems, as information transmission (in the form of nervous spikes) and their consequent processing affects energy consumption. Thus, biological systems tend to minimize the number of spikes (events) to be transmitted and processed for a given computational task. This principle is what neuromorphic computing intends to pursue. Fig. 3 shows an illustrative example of this efficient frame-free event-driven information encoding [⁴⁶]. In Fig. 3(a) we see a poker card deck being browsed at natural speed, recorded with a DVS, and played back at real-time seed with a reconstructed frame time of about 20ms. In Fig. 3(b) the same recorded list of events is played back at 77us frame time. In Fig. 3(c) we show the tracked symbol input fed to a spiking convolutional neural network for object recognition, displaying the recognized output symbol. In Fig. 3(d) we show the 4-layer spiking convent structure, and in Fig. 3 (e) we show the {x,y, time} representation of 20ms input and output events occurring during a change of card so that the recognition switches from one symbol to the next in less than 2ms. Note that here the system is composed of both, the sensor and the network executing the recognition. Both working together need less than 2ms. This contrasts dramatically with conventional artificial systems, in which the sensor first needs to acquire two consecutive images (typically 25ms per image) and then process both to capture the change.

Figure 3: (a) Fast speed poker deck browsing: events are collected about every 20ms to build a frame to display on a computer screen. (b) Slow speed playback at 77us per reconstructed frame. (c) Poker symbol tracked and displayed on the right, and recognition output on the left. (d) Event-driven CNN to classify four poker symbols. (e) {x,y,t} display during 20ms showing camera events together with recognition events during a change of card with a recognition of less than 2ms.

Fig. 3 illustrates a simple version of a neuromorphic sensing and processing system. By today, much larger neuromorphic systems, inspired in the same information encoding scheme, have been developed and demonstrated. Some powerful example systems are:

-        The SpiNNaker Platform [⁴⁷], developed partly within the human Brain Project [⁴⁸], is based on an 18 ARM core SpiNNaker chip, 48 of which are assembled into a node PCB, and about 1200 of which are assembled in a set of furnitures each with racks, hosting all together about 1 million ARM cores. This system is capable of emulating 1 billion neurons in real time. An updated SpiNNaker chip has already been developed performing about 10x in efficiency, neuron emulation capability, and event traffic handling, while keeping similar power consumption.

-        The BrainScales Platform [⁴⁹], also developed during the the Human Brain Project [⁴⁷], implements physical silicon neurons fabricated on full silicon 8-inch wafers, and interconnecting 20 of these wafers in a cabinet, together with 48 FPGA based communication modules. It implements accelerated time computations with respect to real time (about 10,000x), with Spike-Timing-Dependent plastic synapses. Each wafer can host about 200k neurons and 44 million synapses.

-        The IBM TrueNorth chip [⁵⁰] could host 1 million very simple neurons, or be reconfigured to trade-off number of neurons versus neuron model complexity. They were structured into 4096 identical cores, consuming about 63 mW each.

-        Loihi from Intel is probably by today the most advanced neuromorphic chip. In its first version [⁵¹], fabricated in 14nm, it contains 128 cores, each capable of implementing 1k spiking neuronal units (compartments), and including plastic synapses. More recently, Loihi 2 chip was introduced, with up to 1 million neurons per chip, manufactured in Intel 4 technology (7nm). Up to 768 of Loihi chips have been assembled into the Pohoiki Springs system, while operating at less than 500 watts [⁵²].

3.1.3. Challenges and Conclusion

Neuromorphic computing algorithms should be optimum when run on neuromorphic hardware, where events travel and are processing in a fully parallel manner. One of the main challenges in present day neuromorphic computing is to train and execute powerful computing systems directly on neuromorphic hardware. Traditionally, neuromorphic computing problems were mapped to more traditional deep neural networks to obtain their parameters through back-propagation based training [⁵³], which would then be mapped to their neuromorphic/spiking counterpart⁴⁶]. However, these transformations always resulted in a loss of performance. By today there are many proposals of training directly in the spiking domain, combining variants of Spike-Timing-Dependent plasticity rules, with surrogate training techniques that adapt backpropagation to spiking systems, and tested on either on fully connected or convolution based deep spiking neural networks. For an updated review readers are referred to [⁵⁴].

On the other hand, it remains to see whether novel nanomaterial devices, such as memristors, can provide truly giga-scale compact chips with billions of neurons on a single chip and self-learning algorithms. Some initial demonstrations of single [⁵⁵] or multi-core systems [⁵⁶] exploiting a nano-scale memristor combined with a selector transistor as synaptic element have been reported, with highly promising outlooks once synapse elements could be provided as pure nanoscale devices while stacking multiple layers of synapse fabrics together with other nano-scale neurons [⁵⁷]. In the end, the success of neuromorphic computing will rely on the optimum combined progress in neuromorphic hardware, most probably exploiting emerging nano-scale devices massively, in an event- and data-driven information and energy-efficient processing methodologies, and finally in providing efficient, resilient, and quick learning methodologies for mapping real-world applications into the available hardware and computational neuromorphic substrates.

3.2 Different neuromorphic technologies and state of the art

Sabina Spiga

3.2.1. Status

The research field of neuromorphic computing has been growing significantly over the last three decades, following the pioneering research at Caltech (USA) by Carved Mead & co-workers⁵, and it is currently attracting the interest of a wide and interdisciplinary community from device, circuits and systems to neuroscience, biology, computer science, materials and physics. Within this framework, the developed neuromorphic hardware technologies span from fully CMOS-based systems⁵⁸⁵⁹ to solutions exploiting the use of charge-based or resistive non-volatile memory technologies⁶⁰⁶¹⁶², and to emerging memristive device concepts and novel materials^63646566. Fig. 4 reports a schematic (and non-exhaustive) evolution of the main technologies of interest. A common feature of these approaches is to take inspiration from the brain computation, by co-locating memory and processing (in-memory computing-IMC approach), and then overcoming the von-Neumann architecture. Hardware artificial neural networks (ANN) can implement IMC computing and provides an efficient physical substrate for machine learning algorithms and artificial intelligence (AI). On the other side, spiking neural networks (SNN), encoding and processing  information using spikes, hold great promise for applications requiring always-on real-time processing of sensory signals, for example in edge computing, personalized medicine and Internet of Things.

Figure 4: Schematic  evolution of the main hardware technologies of interest for neuromorphic computing (the indicates decades represent only a time frame). Triangular symbols mark the refernce period for early stage studies or starting interest in the technology development. From bottom to top of the figure, the listed technologies are today at higher maturity level and more advanced at system integration level. 

In terms of the maturity of neuromorphic technologies, we can discuss three main bocks.

(i) Current large-scale hardware neuromorphic computing systems are fully CMOS-based and exploit digital or analogue/mixed-signal technologies. Examples of fabricated chips are the IBM TrueNorth, Intel Loihi, Tianjic, ODIN, and others as discussed in these review papers ⁵⁸⁵⁹. In these systems, the neuron and synapse functionalities are emulated by using circuit blocks based on CMOS transistors, capacitors and volatile SRAM memory. The scientific community is now exploiting these chips to implement novel algorithms for AI applications.

(ii) Non-volatile memory technologies. In the last decade, resistive non-volatile memory (NVM) technologies, such as  Resistive Random Access Memory (RRAM), Phase change memory (PCM), Ferroelectric memory (FeRAM) and ferroelectric transistor (FeFET), and magnetoresistive random access memories (MRAM), have been proposed as possible compact, low power and dynamical elements to implement in hardware the synaptic nodes, replacing SRAMs, or as key element of neuronal block^67,61,68. While these NVMs have been developed over the last twenty years mainly for data storage applications, and introduced in the market, they can be considered emerging technologies in the field of neuromorphic computing and their great potential is still not fully exploited. Over the last 10 years, novel concepts for computing, based on hybrid CMOS/non-volatile resistive memory circuits and chips⁵⁶,  have been proposed in the literature. In parallel, also more conventional charge-based non-volatile memories such as FLASH and NRAM are currently being investigated for IMC since they are mature technologies. Finally, it is worth mentioning the emerging memory technologies that are attracting increasing interest in the field of IMC and neuromorphic computing, namely the ferroelectric tunnel junction (FTJ)⁶⁹ and the 3-terminal electrochemical random access memory (ECRAM)⁷⁰.

ii) Advanced memristive materials, devices and novel computation concepts that are currently investigated include 2D materials, organic material, perovskite, nanotubes, self-assembled nano-objects and nanowire networks, advanced device concepts in the field of spintronics (domain wall, race-trace memory, skyrmions), devices based on metal-insulator transition (for instance VO₂-based devices), and volatile memristors.^667165 7273. These technologies are currently proof of concepts at a single device level and circuit blocks connecting a reduced number of devices. The computing system is sometimes demonstrated with a mixed hardware/software approach, where the measured device characteristics are used to simulate large systems. Finally, it is worth mentioning the increasing interest in architectures that can exploit photonics components for computing, towards the building of neuromorphic photonics processors taking advantage of the silicon photonic platforms and co-integration with novel optical memory devices and advanced materials such as phase-change materials  ⁷⁴⁷⁵.

Fig.5 shows schematically examples of the material systems currently most investigated in various approaches and technologies for neuromorphic computing.

Figure 5: Examples of materials systems currently employed in memristive technologies. The list of materials is not exhastive andinclude only some of the most used ones. For the NVM devices(top line), the main active material is indicated, but each device includes also various types of material electrodes depending on the technology. 

3.2.2. Challenges

The current and future challenges can be considered at various levels.

(i) For large-scale neuromorphic processors, the progress of CMOS-based technologies and their scaling still provide room to advance the research field. The main challenges are at the architecture and algorithms level. On the other side, most NVM memories (RRAM, PCM, FeRAM, FeFET and MRAM) have been already integrated with CMOS at scaled technological nodes and large integration density, and hold interesting properties (depending on the specific technology) such as small size, scalability, possible easy integration also in 3D array stacking, low programming energy, multilevel programming capability.   Therefore, it is expected that NVM technologies will play an increasing role in future IMC chips or neuromorphic processors, by enabling energy-efficient computation. Prototype IMC chips have been reported in the literature⁵⁶⁷⁶, as well as innovative circuits for SNN implementing advanced learning rules to compute with dynamics⁷⁷⁷⁸.

On the other side, It is worth mentioning that the NVM technologies exhibit several device-level non-idealities., as discussed in more detail in the following sections of this roadmap. As examples and non-exhaustive list, we can mention: nonlinearity and stochasticity in conductance update vs a number of pulses at a fixed voltage (PCM, RRAM, FeRAM), asymmetry (RRAM) in the bidirectional tuning of conductance, conductance drift (PCM) or broadening of the resistance distribution (RRAM) after programming, device-to-device and cycle-to-cycle variability of the programmed states, low resolution due to the limited number of programmable levels (up to 8 or 16 are demonstrated for RRAM and PCM at array level), restricted memory window (MRAM) or limited endurance (general issue except than for MRAM), relative high conduction also in the OFF state.   All these aspects can impact the neural network accuracy and reliability, although proper algorithms/architecture can take advantage from stochasticity or asymmetry of conductance tuning ⁷⁹. Therefore, a careful co-design of hardware and algorithms is required together with an improvement of circuit design and/or programming device strategies to fully exploit NVMs in combination with CMOS and in large systems. Further discussion on specific challenges and  possible specific applications of the listed technologies will be discussed in the sections 5.1 - 5.4 of this roadmap, while a more deep view on application scenario is reported in section 7.

ii) Regarding the plethora of emerging materials/devices and novel concepts proposed for neuromorphic computing (beyond the ones discussed in the previous point, see some examples in section 5.4-5.6 of the roadmap), the main challenge is that they are mostly demonstrated at the single device level or in early stage proof of concepts in small array/large device size, then extending their implementation in ANN or SNN only at simulation level. To leaverage these concepts at higher TRL, it is necessary to prove that the device characteristics are reproducible and scalable, the working principle well understood, and to provide more advanced characterizations on several down-scaled devices, and fianlly to close the current gap between laboratory exploration of single material/devices and integration in array or circuits. Another challenge is to address more into details how to exploit the nanodevcies peculiarities, such as dynamic or stochastic behavior, to implement in hardware more complex bio-inspired functionalities or even to perform radically new computation paradigms. Indeed, while the more standard technologies (CMOS, Flash,SRAM) can also be used in hardware neural network to implement complex functions,  this is possible only at the high cost of increased circuit complexity. To give an example, the required dynamic to reproduce the synaptic or neuronal fucntionality in SNN is implemented at circuit level and/or using large area capacitors which are not easy scalable in view of large systems. One possible approach is to exploit the emerging memristve technologies and their properties (variability, stochasticity, non-idealities) to implement complex functions with more compact and low-power devices. One example is the use of resistivity drift in PCM (usually an unwanted characteristics for IMC or storage applications) to implement advanced learning rules in hardware SNN⁸⁰. Another example (discussed in section 7.1) is to use the inherent variability and stochasticity of some nanodevices to build efficient random number generators ( for data security applications) and stochastic computing models. Overall, this scenario point out a long-term development research, likely up to ten years or more, to close the gap between these novel concepts and real industrial applications.

3.2.3. Potential Solutions

To pursue advances in the development of neuromorphic hardware chips, it is necessary to develop a common framework to compare and benchmark different approaches, also in view of some metrics such as computing density, energy efficiency, computing accuracy, learning algorithms, theoretical framework as well as target possible killer applications that might significantly benefit from neuro-inspired chips. Within this framework, materials strategies can be still relevant to address some of the outlined challenges for NVMs, but materials need to be co-developed together with a demonstration of a device at the scaled node and array level. An important strategy for future is also the possibility to substitute current mainstream materials with green materials or to identify fabrication processes more sustainable in term of cost and environmental impact, without compromising the hardware functionality. Moreover, other important aspects include the development of hardware architecture that can lead to the integration of several devices, and exploiting a large connectivity among them; the implementation of efficient algorithms supporting online learning, also on different time scales as in biological systems; and to address the low power analysis of large amount of data also for internet of things and edge devices. Overall, it is necessary and holistic view which includes the materials/device/architecture/algorithms co-design to develop large-scaled neuromorphic chip.

3.2.4 Concluding remarks

The development of advanced neuromorphic hardware than can support efficiently AI applications is becoming more and more important. Despite the several prototypes and results presented in literature, neuro-inspired chips are still only an early stage of development and there is plenty of room for further development. Many mature NVM devices are definitely candidates to became a future mainstream technology for large scale neuromorphic processor that can outperform the current platform based only on CMOS circuits. In the long term, it is also necessary to close the gap between emerging materials and concepts, currently demonstrated only at proofs of concepts, and their possible integration in functional systems.  Materials research and understading of physical principles enabling novel functioalities are an important parts fo this scenario.

3.3. Possible future computational primitives for neuromorphic computing

Sergey Savel’ev, Alexander Balanov

The core idea of neuromorphic computing to develop and design computational systems mimicking electrochemical activities in brain cortex is currently booming, embracing areas of deep physical neural networks ⁸¹, classical and quantum reservoir computing ⁸²,⁸³, oscillator-based computing  ⁸⁴, and spiking networks ⁸⁵  among many other concepts ⁸⁶. These computational paradigms imply new ways for information processing and storage different from conventional computing and therefore require elementary base and primitives which often involve unusual novel physical principles ⁸⁷.

Presently, memristors - electronic switchers with memory - and their circuits demonstrate great potential for application in the primitives for future neuromorphic computing systems. In particular, different types of volatile and non-volatile memristors can serve as artificial neurons and synapses, respectively, which facilitate the transfer, storage, and processing of information ⁸⁸. For example, volatile Mott memristors  ⁸⁹ can work as an electric oscillator with either regular or chaotic dynamics ⁷, while memristors with filament-formation ⁹⁰ demonstrate tunable stochasticity ⁹¹, ⁹² allows designing neuromorphic circuits with different degree of plasticity, chaoticity, and stochasticity to address diverse computational aims in mimicking dynamics of different neuron populations. Furthermore, a crossbar of non-volatile memristors (servicing to memorise training) attached to volatile memristors (working as readouts) enables the design of AI hardware with unsupervised learning capability ⁹³. Thus, combining memristive circuits with different functionalities paves the way to building a wide range of in-memory computational blocks for a broad spectrum of artificial neural networks (ANNs) starting from deep learning accelerators to spiking neuron networks ⁹⁴.

A rapidly developing class of volatile memristive elements ⁹⁵ has been shown to demonstrate a rich spectrum of versatile dynamical patterns ⁹⁶,^7 97, which makes them suitable for the realisation of a range of neuroscience-motivated AI concepts ⁹⁸, ⁹⁹, ¹⁰⁰. For instance, the ANNs based on volatile memristors can go well beyond usual oscillator-based computing ⁸⁴ or spiking neural networks ⁸⁵. They rely on manipulating information by utilising complexity in dynamical regimes that offer a novel computational framework ⁹⁸, ⁹⁹ with cognitive abilities closer to biological brains. There is a specific emphasis on using dynamical behaviours of memristors, instead of only static behaviours. ¹⁰¹

Remarkably memristive elements can be realised not only in electronic devices but also within spintronics or photonics frameworks, which have their own advantages compared to electronics. Therefore, hybridised design promises great benefits in the further development of neuromorphic primitives. For example, a combination of memristive chipsets with spintronic or/and photonic components can potentially create AI hardware with enhanced parallelism offered by optical devices operating simultaneously at many frequencies (e.g., optical cavity eigenfrequencies) ¹⁰², energy-efficient magnetic non-volatile memory and flexible memristive spiking network architecture. An important step in the realisation of this approach is the development of interface technologies for bringing electronic, photonic and spintronic technologies together. A possible example is the spintronic memristor ¹⁰³, ¹⁰⁴, ⁶⁵, where the transformation of magnetic structure influences the resistance of the system. Control of resistance through magnetic structures by either current or voltage and vice versa offers a possibility to combine the same chipset spintronic both components ¹⁰⁵ affecting the magnetic structure of the spintronic memristor and memristive components influencing current through the spintronic memristor. Therefore, this promotes crosstalk between the electric and magnetic subsystems of the same device. An interface between neuromorphic optical and electronic sub-systems of a hybrid device could be realised using optically controlled electronic memristive systems ¹⁰⁶, thus, paving a path towards neuromorphic opto-electronic systems ¹⁰⁷.

The conventional ANNs with a large number of connections require training to is is less efficient in the task requiring frequent retraining for ‘’moving target” problems, for example in recognition of characteristics changing in time. A potential solution for such tasks is to implement filtering or pre-processing data by a “reservoir” ⁸², usually consisting of neuron units connected by fixed weights. The reservoir is assisted by a small readout ANN, which requires much less data for training thus removing significant retraining burden. Recently, an important evolution has taken place in the development reservoir computing systems, where the function of the reservoir is realised by photon, phonon or/and magnon mode mixing in spintronic ¹⁰⁸, ¹⁰⁹, and photonics ¹¹⁰ devices. Substitution of the interaction of many artificial neurons by wave processes resembles neural wave computation in the visual cortex ⁹⁹ and promotes miniaturisation, robustness, and energy efficiency of the reservoirs (neuromorphic accelerators) which in future could become an additional class of primitive, especially in neuromorphic computational systems dealing with temporal or sequential data processing ¹¹¹. In AI training, it has also been shown that memristive matrix multiplication hardware can enable noisy local learning algorithms, which perform training at the edge with significant energy efficiencies compared to graphics processing units. ¹¹²

Finally, we briefly outline another exciting perspective constituted by a combination of quantum and neuromorphic technologies ¹¹³. Currently, quantum AI ¹¹⁴ attracts significant attention by increasingly competing with more traditional quantum computing. One of the most promising quantum AI paradigms is quantum reservoir computing, ¹¹⁵ which offers not only much larger state space than classical reservoir computing but also essentially-nonclassical quantum feedback on the reservoir via measurements. A quantum reservoir built from quantum memristors ¹¹⁶, ¹¹⁷, could significantly gain quantum AI efficiency as it can readily be integrated with existing quantum and classical AI devices and also lead to an ‘’exponential growth” ¹¹⁸ in the performance of “reservoirs” with the possibility of relaxing requirements on decoherence comparing to traditional quantum computing.

The above trends and directions in the development of the primitives for neuromorphic computing are obviously only a slice of exciting future AI hardware technology. Even though we recognise that our choice is subjective, we hope that the outlined systems should provide a flavour of future computational hardware, which should be based on reconfigurable life-mimicking devices utilising different physical principles in combination with novel mathematical cognitive paradigms ¹¹⁹, ¹²⁰, ¹²¹, ¹²².

4. Mature Technologies

4.1 SRAM

Nitin Chawla, Giuseppe Desoli

4.1.1 Status

SRAM-based computing in memory (CIM) or in-memory computing, is seen as a mature and widely available technology for accelerating matrix and vector calculations in deep learning applications, yet many technology driven optimizations are still possible. To make CIM more compatible, researchers have been exploring ways to improve the design of the bit-cell (Fig. 6), which is the basic unit of memory. This has led to the development of high-end SRAM chips with large capacities, such as 107, 128 Mb, and 256 Mb SRAM chips at 10 nm, 7 nm, 5 nm, and 4 nm ^123124125126. These large SRAM capacities help to reduce the need for off-chip DRAM access. However, in more cost-sensitive applications, such as embedded systems and consumer products, modifying the bit-cell design can be too costly and may limit the ability to easily transfer the technology to different manufacturing nodes.

A key difference exists between analog and digital SRAM CIM. Analog CIM has been heavily studied using capacitive or resistive sharing techniques to maximize row parallelism ¹²⁷, but this comes at the cost of inaccuracies and loss of resolution due to variations in devices across PVT and the limitations of SNR and dynamic range in ADC/readout circuits. Impacts of device variations for different kinds of devices:

• Resistive devices like PCM or RRAM experience a variation in the resistive values across the nominal behavior which can vary based on process and for a case of +/- 10-20% change in resistance value there will be corresponding change of current values which are then input to the read out circuits and hence this will impact the quantization step of read out circuits hence impacting the SNR which will then need a higher dynamic range to compensate for the same. Temperature behavior for resistors also needs to be taken care in the noise margin.

• MOS devices:  These devices can vary in their performance (threshold voltage) due to:

1. Global lot positioning like slow, typical, fast which can again vary around +/- 20% which can be less or more based on technology and voltage of operation. This is a deterministic shift.

2. Local variation: within the same lot there are device to device variations which are random in nature and need statistical analysis based on capacity in use to analyze the impact of variations. These impact the SNR and Quantization like in case of resistive devices and will need higher dynamic range to compensate for the loss in accuracy.

Analog SRAM CIM solutions often use large logic bit cells and aggressive reduction in ADC/readout bit width, resulting in low memory density and computing inaccuracies, making it difficult to use in situations where functional safety, low-cost testing, and system scalability are required. On the other hand, digital CIM offers a fast path for the next generation of Neural Processing systems due to its deterministic behavior and compatibility with technology scaling rules.

Researchers have improved the SRAM-based CIM’s performance by modifying the SRAM bitcell structure and developing auxiliary peripheral circuits. They proposed read-write isolation cells to prevent storage damage and transposable cells to overcome storage arrangement limitations. Peripheral circuits, such as DAC, redundant reference columns, and multiplexed ADC, were proposed to convert between analog and digital signals.  The memory cell takes up most of the SRAM area in the core module of a standard SRAM cut. However the complexity of the additional operations performed in the memory unit, poses additional problems to utilize the memory cells to their full potential. Researchers have explored various trade-offs to implement the necessary computational functionality while preserving density, power, and, last but not least, minimizing the additional cost associated with bitcell modifications required for requalification when deployed in standard design flows.  Most system-on-chips (SoCs) use standard 6T structures due to their high robustness and access speed and to minimize area overhead. The 6T storage cell is made up of two PMOSs and four NMOSs to store data stably. To perform CIM using the conventional 6T SRAM cell, operands are represented by the WL voltage and storage node data, and processing results are reflected by the voltage difference between BL and BLB.

Figure 6: Standard SRAMs bitcells are usually designed with 6 or 8 transistors

The figure above shows the conventional 6T and 8T bitcells which form the basic building block of SRAM design. The 8T bitcell is made out of a conventional 6T and read port which allows read and write in parallel. These bitcells were never designed for parallel access across rows and this poses one of the main challenges for enabling analog SRAM CIM.

Dual-split 6T cell with double separation have been proposed ^128,129, allowing for more sophisticated functions due to the separated WL and GND, which can use different voltages to represent various types of information. ¹³⁰  proposed a 4+2T SRAM cell to decouple data from the read path. The read is akin to that of the standard 6T SRAM, writes instead, use the N-well as the WWL and two PMOS source as the WBL and WBLB. In computational mode, different voltages on the WL and storage node encode the operands.

In general, CIM adopting the 6T bitcell structure are unable to efficiently perform computing operations and may not fully meet the requirements of future CIM architectures. Hence, many studies on CIM have modified the 6T structure because using the 6T standard cell directly poses a reliability challenge as the contents of the bitcells get effectively shorted if accessed in parallel on the same bitline. This means special handling on the wordline voltage is required which adds lot of complexity and limits the dynamic range. Further the variability and linearity of devices become very difficult to control if when limiting the device operation to reduced voltage levels due to these reliability constraints, impacting the overall energy efficiency of the solution ^131,132.

For practical applications, and specifically for AI ones, it’s important to evaluate the end to end algoritmic accuracy vs. the key metrics, to this end, recent research ^{133134135136 137} has suggested various analytical models to examine the balance between the costs (accuracy) and benefits (primarily energy efficiency and performance) of digital versus analog SRAM CIM. This is based on the idea that many machine and deep learning algorithms can tolerate some degree of computational errors, and that there are methods such as retraining and fine-tuning as well as hardware-aware training to address these errors.

The implementation of Neural Processing Units incorporating CIM components for large-scale deep neural networks (DNNs) presents significant difficulties, CIM macro can incur substantial column current magnitudes, which can result in power delivery difficulties and sensing malfunctions. Furthermore, the utilization of analog domain operations necessitates the incorporation of ADCs and DACs, which consume a significant amount of area and energy resources. Further to this, the pitch matching of ADC with SRAM  bitcell also poses a big challenge for array and ADC interface. It is clear that the realization of the full potential of SRAM-CIM necessitates the development of innovative and sophisticated techniques.

4.1.2 Challenges and Potential Solutions

In deep learning, convolutional kernels and other types of kernels rely heavily on matrix/vector and matrix/matrix multiplications (MVM). These operations are computationally expensive and involve dot product operations between activation and kernel values. In-memory multiplication in CIM macro devices can be classified into three primary categories: current-based, charge-sharing-based for analog computation, and one for all-digital. All-digital CIMs exhibit the same level of precision as purely digital ASIC implementations. Various implementation topologies ranging from bit-serial to all parallel arithmetic implementations have been proposed for digital CIM solutions. Digital CIMs like in research work ¹³⁸ represent a modified logic bit-cell to support element-wise multiplication followed by a digital accumulation tree sandwiched within the SRAM array. The solution improves on energy efficiency by reducing data movement alongside the efficiency benefits of a custom-built MAC pipeline with improved levels of parallelism over traditional Digital NPU’s for example as in Fig. 7 ¹³⁹. The Digital CIM implementations also have a wide voltage and frequency dynamic range allowing runtime reconfigurability between the competing TOPS/W and TOPS/mm2 performance criteria. The operating range and mission profiles of these architectures can also be extended by leveraging read-and-write assist schemes as is commonly done for ultra-low voltage SRAM design. The digital CIM solution’s energy efficiency depends on the operand precision and due to the deterministic precision and bit true computation nature, it begins to decline as we increase the operand precision.

Figure 7: Digital CIM memory macro with 8T bitcells and embedded digital logic

Current-based CIMs as represented in one of the early research works ¹⁴⁰, implements a WL DAC driving a multi-level feature input with multiple rows active in parallel. The results of the element-wise multiplications of all the parallel rows are accumulated as current on the bit lines of the CIM macro that terminates in a current-based readout/ADC The current accumulation on the bit line essentially implements a reduction operation limited by the SNR of the readout circuit. Current-based CIMs as presented in this work suffers from significant degradation in accuracy due to bit cell variabilities and nonlinearities of the WL DAC while the throughput is limited by the readout circuits. ¹⁴¹ implements a variation with a CIM using a 6 T-derived bit cell with a PWM WL modulation and a focus on storing and computing multibit weights per column. The modulation scheme uses binary weighted pulse duration based on the index of the bit-cell in the column effectively encoding the multi-bit weight in the column to impact the value of the global bit line. The multiplication is effectively done in the periphery of this CIM using a switched capacitor circuit.Bit-cell variation and non-linearity like in the previous case significantly limit the accuracy of this implementation thus restricting the industrialization potential of these current-based CIM solutions. The work in ¹⁴² tries to address the limitations of the above analog CIM techniques and implements a charge-based CIM by using a modified SRAM logic bit-cell that performs an element-wise binary multiplication(XNOR) and transfers the results to a small capacitor. Multiple rows operating in parallel as key to the energy efficiency of these CIM topologies. In this work, the element-wise multiplication result is transferred as a charge to the global bit line followed by a voltage-based readout. The inherent implementation benefits from the fact that capacitors suffer less from process variability and present fairly linear transfer characteristics. This architecture however like other Analog CIM is impacted by dynamic range compression due to the limited SNR regime of the readout at the end of the column. ¹⁴³ extends this approach to support multi-bit implementations using a bit-sliced architecture. The multi-bit weights are mapped to different columns while the feature data is essentially transferred as 1-bit serial data on parallel word lines and each column performs a binary multiplication followed by accumulation on the respective bit lines. The near-memory all-digital recombination unit in this approach performs the shift and scale operations based on the column index to recreate the results of the multi-bit MAC operation. The approach is flexible to support asymmetric features and weight precision and can be made reconfigurable to support different features and weight precisions on the fly. This however still suffers from the same SNR constraints as each column operation is compressing the dynamic range and is limited by the peak dynamic range of the readout ADC. The ADC in most of these schemes is mostly shared across multiple columns thus making it a critical design component in determining the throughput of such CIM architectures. The specific bit-sliced approach has impressive TOPS/W numbers for the lower weights and activation precision regime but starts to taper off due to the quadratic increase in the computation energy with increasing weight and activation precision. ¹⁴⁴ instantiates multiple of these CIM macros to demonstrate a system-level approach connecting these CIM macros with a flexible interconnect and adding digital SIMD and scalar arithmetic units to support real-world Neural Network execution. This specific work due to the limited readout speed of the CIM macros and the overhead of the other digital units suffers from a moderate TOPS/mm2 number for the full solution but presents impressive TOPS/W numbers, especially at the lower precision regime. The research work ¹⁴⁵ represents another effort with a system-level solution of a hybrid NPU comprising analog CIM units and traditional digital accelerator blocks. The work leverages a low-precision (2-bit) Analog CIM macro coupled with a traditional 8-bit digital MAC accelerator. The two orders of magnitude difference in energy efficiency between the 8-bit digital MAC engine and the 2-bit Analog CIM macro can be leveraged by mapping different layers to the appropriate computation engine but needs careful articulation of mapping algorithms with the precision constraints of the Analog CIM’s while keeping the overhead of the write refresh and other digital vector/scalar operators low. This to some expect is a tradeoff between a very specialized use case as opposed to a general-purpose NN accelerator.

4.1.3 Conclusions

Analog CIM solutions based on charge-based CIMs display a lower degree of variability when compared to current-based CIMs, due to variability in the technologies employed for capacitors and threshold voltage effects. Additionally, charge-based CIMs are able to activate a greater number of word lines per cycle and thus achieving higher amounts of row parallelism. However, both current-based and charge-based CIMs are limited in terms of accuracy and the equivalent bit precision of the dynamic range of the accumulation. Selecting an appropriate ADC bit-precision and MVM parallelism is a challenging task that requires balancing accuracy and power consumption. Measurements and empirical evidence suggest that an increase in the accumulation values is correlated with a higher degree of variability. However, it is important to note that such high values are relatively rare in practical neural network models, as shown by the statistical distribution of activation data and resultant accumulation outcomes. This characteristic along with noise-aware training can be leveraged to optimize the precision and throughput demands of the analog-to-digital converter, thereby improving the FOM of these Analog CIM techniques. The research on noise-aware training in the state of the art is limited to academic works on relatively small neural networks and data sets. This for an industrial deployment still needs to mature and demonstrate the scalability to larger models and data sets.

All-digital CIMs provide a deterministic and scalable path to intercept the implementation of NPUs by bringing an order or more of gain vs traditional all-digital NPUs. Digital CIM solutions provide excellent scaling for area and energy efficiency as we move towards more advanced CMOS nodes with a wide operating voltage and frequency range tunability while still maintaining a general purpose and application-agnostic view of embedded neural network acceleration at the edge.

On the other hand for applications that can handle approximate computing, Analog SRAM CIM-based solutions provide a much-increased level of computation parallelism and energy efficiency while still operating in an SNR-limited regime. The impact of dynamic range compression and readout throughput are key algorithmic and design tradeoffs while designing an Analog CIM solution which tries to operate in a much more restricted voltage and frequency regime as opposed to a Digital CIM solution. Given the application choices being more vertically defined as opposed to general purpose is also a deciding factor in choosing an Analog SRAM CIM-based solution as opposed to  Digital CIM solutions. In conclusion, due to the rapid industrialization potential of SRAM-based CIM solutions and the opportunity of exploiting the duality of these CIM instances to serve as SRAM capacity to support the system in other operating modes, there are enough reasons to remain invested in SRAM CIM. The scope to improve both digital and analog SRAM CIM remains very high, both at the design and technology level, to exploit the best gains out of these two solutions which in the future can also be combined to form a hybrid solution serving multiple modalities of Neural Network execution at the Edge.

4.2  Flash memories

Gerardo Malavena, Christian Monzio Compagnoni 

4.2.1 Status 

Thanks to a relentless expansion in all the application fields of electronics since their conception in the 80’s of the 20th century, Flash memories became ubiquitous nonvolatile storage media in everyday life and a source of market revenues exceeding $60B in 2021. The origin of this success can be traced back to their capability to solve the trade-off against cost, performance and reliability in data storage much better than any other technology. Multiple solutions to that trade-off, besides, were devised through different design strategies that, in the end, allowed Flash memories to target a great variety of applications in the best possible way. Among these different design strategies, the two leading to the so-called NOR Flash memories ¹⁴⁶ and NAND Flash memories ¹⁴⁷ became by far the most important.

As in all Flash memory designs, NOR and NAND Flash memories store information in memory transistors arranged in an array whose operation relies on an initialization, or erase, step performed in a flash on a large number of devices simultaneously. In particular, the erase step moves the threshold-voltage (V_T) of all the memory transistors in a block/sector of the array to a low value. From that initial condition, data are stored through program steps performed in parallel on a much smaller subset of memory transistors and raising their V_T to one or more predefined levels. This working scheme of the array allows to minimize the number of service elements needed for information storage and, in the end, is at the basis of the high integration density, high performance and high reliability of Flash memories. Starting from it, the structure of the memory transistors, the architectural connections among them to form the memory array, the array segmentation in the memory chip, the physical processes exploited for the erase and program steps, and many other aspects are markedly different in NOR and NAND Flash memories.

NOR Flash memories follow a design strategy targeting the minimization of the random access time to the stored data, reaching latencies as short as a few tens of nanoseconds. A strong array segmentation is then adopted to reduce the delay time of the word-lines (WLs) and bit-lines (BLs) driving the memory transistors. As depicted in Fig.8(a), moreover, the memory transistors are independently connected to the WLs, BLs and source lines (SLs) of the array to simplify and speed up the sequence of steps needed to randomly access the stored data and to allow device operation at relatively high currents (currents in the microAmpere scale are typical to sense the data stored in the memory transistors). Fast random access is also achieved through a very robust raw array reliability, with no or limited adoption of error correction codes (ECCs). This design strategy, on the other hand, does not make NOR Flash memories the most convenient solution from the standpoint of the area and, hence, the cost of the memory chip and limit the chip storage capacity to low or medium sizes (up to a few Gbit).

NAND Flash memories rely on a design strategy pointing to the minimization of the data storage cost. Therefore, limited array segmentation is adopted and the memory transistors are in series connection along strings to reduce the area occupancy of the memory chip. Figs.8(b-1) and (b-2) schematically show the arrangement of the memory transistors in a planar and in a vertical (or, 3-D) NAND Flash array, respectively. 3-D arrays represent today the mainstream solution for NAND Flash memories, capable of pushing their bit storage density up to  [?]  15Gbit/mm^2 148, a level unreachable by any other storage technology. Such an achievement was made possible also by the use of multi-bit storage per memory transistor and resulted in memory chips with capacity as high as 1Tbit ¹⁴⁸ . The NAND Flash memory design strategy, on the other hand, makes the random access time to the stored data relatively long (typically, a few tens of microseconds). That is the outcome of time delays of the long WLs and BLs in the microsecond timescale, low sensing currents (tens of nanoAmpere) during data retrieval due to series resistance limitations in the strings, and the need of multi-bit detection per memory transistor. Besides, array reliability relies on powerful ECCs.

Given the successful achievements of Flash memories as nonvolatile storage media for digital data, exploiting them in the emerging neuromorphic-computing landscape appears as a natural expansion of their application fields and is attracting widespread interests. In this landscape, Flash memories may work not only as storage elements for the parameters of artificial neural networks (ANNs), but also as active computing elements to overcome the von Neumann bottleneck of conventional computing platforms. The latter may represent, of course, the most innovative and disruptive application of Flash memories in the years to come. At the same time, the use of Flash memories as active computing elements may boost the performance, enhance the power-efficiency and reduce the cost of ANNs, making their bright future even brighter. In this context, relevant research has been focusing on employing Flash memory arrays as artificial synaptic arrays in hardware ANNs and as hardware accelerators for the vector-by-matrix multiplication (VMM), representing the most common operation in ANN. Quite promising results have already been reported in the field, through either NOR ^149,150 or NAND ^{151,152,153,154,155,156} Flash memories. In these proofs of concept, different encoding schemes for the inputs (e.g., voltage amplitude or pulse width modulation, with signals on the BLs or WLs of the memory array) and different working regimes of the memory transistors have been successfully explored.   Interested readers may go through the references provided in this section for a detailed description of the most relevant schemes proposed so far to operate a Flash array as a computing element.

4.2.2 Challenges and potential solutions

In spite of the encouraging proofs of concept already reported, the path leading to Flash memory-based ANNs still appears long and full of challenges. The latter can be classified in the following categories:

Challenges arising from changes in the design strategy of the array

As previously mentioned, the success of Flash memories as nonvolatile storage media for digital data arises from precise design strategies. Modifying those strategies to meet the requests of ANNs may deeply impact the figures of merit of the technology and should be carefully done. For instance, ANN topologies requiring to decrease the segmentation of NOR Flash arrays may worsen their performance in terms of working speed. Increasing the segmentation of NAND Flash arrays to meet possible ANN topology constraints or to enhance their working speed may significantly worsen their cost per memory transistor.

The cost per memory transistor of Flash memories, besides, is strictly related to the array capacity. Modifying the latter or not exploiting it all through the ANN topology may reduce the cost effectiveness of the technology. In this regard, the very different capacities of NOR and NAND Flash arrays make the former suitable for small/medium size ANNs (less than 1 Giga parameters) and the latter suitable for large size ANNs (more than 1 Giga parameters). The organization of the memory transistors into strings in NAND Flash arrays represents an additional degree of complexity for the exploitation of their full capacity in ANNs. In fact, the number of memory transistors per string is the outcome of technology limitations and cost minimization and, therefore, cannot be freely modified. Exploiting all the memory transistors per string, then, necessarily sets some constraints on the ANN topology (number of hidden layers, number of neurons,…), which, of course, should be compatible with the required ANN performance.

Another important aspect to consider is that the accurate calibration of the V_T of the memory transistors needed by high-performance and reliable ANNs may not be compatible with the block/sector erase scheme representing a cornerstone of all the design strategies of Flash memories. Solutions to carry out the erase step on single memory transistors are then to be devised. These solutions may require a change of the array design as in ^150,149 or new physical processes and biasing schemes of the array lines to accomplish the erase step as in ^157,158,159. All of these approaches, however, necessarily impact relevant aspects of the technology, affecting its cost, performance or reliability, and should be carefully evaluated.

The change of the typical working current of the memory transistors when exploiting Flash memories for ANN applications is another critical point to address. In fact, reducing the working current of the memory transistors may make it more affected by noise and time instabilities. Increasing it too much, on the other hand, may raise issues related to the parasitic resistances of the BLs, SLs, and, in the case of NAND Flash arrays, of the unselected cells in the strings.

Challenges arising from array reliability

 Flash memories are highly-reliable nonvolatile storage media for digital data. That, however, does not assure that they can satisfactorily meet the reliability requirements needed to operate as computing elements for ANN applications. Especially in the case of NAND Flash memories, in fact, array reliability in digital applications is achieved through massive use of ECCs and a variety of smart system-level stratagems to take under control issues such as electrostatic interference between neighboring memory transistors, lateral migration of the stored charge along the charge-trap storage layer of the strings, degradation of memory transistors after program/erase cycles, and so on. All of that can hardly be exploited to assure the reliable operation of Flash arrays as computing elements. Besides, the requirements on the accuracy of the placement and the stability over time of the V_T of the memory transistors when using Flash arrays as computing elements may be more severe than in the case of digital data storage. The possibility to satisfy those requirements in the presence of the well-known constraints to the reliability of all Flash memory designs ^147,160,161 is yet to be fully demonstrated. In this context, periodic recalibration of the V_T of the memory transistors and on-chip learning ¹⁵⁶ may mitigate the array reliability issues.  

Challenges arising from the peripheral circuitry of the array

As in the case of Flash memory chips for nonvolatile storage of digital data, the peripheral circuitry of Flash memory arrays used as computing elements for ANNs should not introduce severe burdens on the chip area, cost, power efficiency and reliability. In the latter case, this aspect is particularly critical due to the need to integrate on the chip not only the circuitry to address the memory transistors in the array and to carry out operations on them, but also, for instance, to switch between the digital and the analog domain in VMM accelerators or to implement artificial neurons in hardware ANNs. Along with effective design solutions at the circuit level ¹⁵¹, process solutions such as CMOS-Under-Array integration ¹⁴⁸  or heterogeneous integration schemes ¹⁵³  should be exploited for successful technology development.

4.2.3 Conclusion

Flash memories may play a key role in the neuromorphic-computing landscape. Expanding their fields of application, they can be the elective storage media for ANN parameters. But they can also be active computing media for high-performance, power-efficient and cost-effective ANNs. To achieve this intriguing goal, relevant challenges must be faced from the standpoint of the array design, reliability and peripheral circuitry. Winning those challenges will be a matter of engineering and scientific breakthroughs and will pave the way to years of unprecedented prosperity for both Flash memories and ANNs.

Figure 8: Schematic for the connection of the memory transistors in: (a) a NOR Flash memory array (a common ground architecture of stacked-gate memory transistors has been assumed);  (b-1) a planar and (b-2) a vertical (3-D) NAND Flash memory arra

5. Emerging Technologies (computing approaches)

Zhongrui Wang, J. Joshua Yang

5.1 Resistive switching and memristor

5.1.1. Status

Resistive switches (often called memristors when device nonlinear dynamics are emphasized) are electrically tunable resistors, of a simple metal-insulator-metal structure. Typically, their resistance changes as a result of redox reactions and ion migrations, driven by electric fields, chemical potentials and temperature  ⁶⁴. There are two types of resistive switches according to the mobile ion species. In many dielectrics, especially transition metal oxides and perovskites, anions such as oxygen ions (or equivalently oxygen vacancies) are relatively mobile and can form a conduction percolation path, leading to the so-called valence change switching. For example, a conical pillar-shaped nanocrystalline filament of Ti₄O₇ Magnéli phase filament was visualized by a transmission electron microscope (TEM) in a Pt/TiO₂/Pt resistive switch¹⁶². On the other hand, the conduction channels can also be created by the redox and migration of cations, which involves the oxidation of an electrochemically active metal such as Ag and Cu, followed by the drift of mobile cations in the solid electrolyte and the nucleation of cations to establish a conducting channel upon reduction. The dynamic switching process of a planar Au/SiO_x:Ag/Au diffusive resistive memory cell was captured by in situ TEM⁹⁰.

Resistive switches provide a hardware solution to address both the von Neumann bottleneck and the slowdown of Moore’s law faced by conventional digital computers. When these resistive switches are grouped into a crossbar array, they can naturally perform vector-matrix multiplication, one of the most expensive and frequent operations in machine learning. The matrix is stored as the conductance of the resistive memory array, where Ohm’s law and Kirchhoff’s current law physically govern the multiplication and summation, respectively⁶⁴. As a result, the data is both stored and processed in the same location. This in-memory computing concept can largely obviate the energy and time overheads incurred by expensive off-chip memory access on conventional digital hardware. In addition, the resistive memory cells are of simple capacitor-like structures, equipping them with excellent scalability and 3D stackability. So far resistive in-memory computing has been used for hardware implementation of deep learning models to handle both unstructured (e.g. general graphs, images, audio and texts) and structured data, as discussed in the following:

General graph: Graph-type data consists of a set of nodes together with a set of edges. The theoretical formulation has been made for graph learning using resistive memory on datasets such as WikiVote^163,164. Experimentally, a resistive memory-based echo state graph neural network has been used to classify graphs in MUTAG and COLLAB datasets as well as nodes in the CORA dataset¹⁶⁵, including few-shot learning of the latter¹⁶⁶.

Images: Images are special graph-type data. Both supervised and unsupervised learning of ordinary images have been experimentally implemented on resistive memory. For supervised learning, offline trained resistive memory, where optimal conductance of memory cells is calculated by digital computers and transferred to resistive memory, is used to classify simple patterns^167,168, MNIST handwritten digits^{169,170,171,172}, CIFAR-10/100 datasets^173,174,175, ImageNet¹⁷⁶; as well as Omniglot one-shot learning dataset¹⁷⁷. In addition to offline training, online training adjusts the conductance of resistive memory in the course of learning, which is more resilident to hardware nonidealities in classifying simple patterns^167,37, Yale face and MNIST datasets^178,179, CIFAR-10 dataset¹⁸⁰, and meta-learning of Omniglot dataset¹⁸¹. Besides supervised learning, unsupervised offline learning with resistive memory is used for sparse coding of images¹⁸² and MNIST image restoration¹⁸³.

Audios and texts: Learning sequence data such as audios and texts have been implemented on resistive memory. Supervised online learning using recurrent nodes has been done on the Johann Sebastian Bach chorales dataset¹⁸⁴. In addition, delayed-feedback systems based on dynamic switching of resistive memory are used for temporal sequence learning, such as spoken number recognition and chaotic series prediction^185,186,187. For offline learning, resistive memory is used for modeling the Penn Treebank dataset¹⁸⁸, Wortschatz Corpora language dataset and Reuters-21578 news dataset¹⁸⁹, as well as Bonn epilepsy electroencephalogram dataset and NIST TI-46 spoken digit dataset^190,191.

Structured data: Despite unstructured data, structured data such as those of a tabular format has been tackled by resistive memory, including supervised classification of the Boston housing dataset on an extreme learning machine¹⁹², K-means clustering of the IRIS dataset and principal component analysis of the breast cancer Wisconsin (diagnostic) dataset^193,194 and correlation detection of quality controlled local climatological database¹⁹⁵.

Figure 9: Summary of emerging memory such as memristors and their capabilities in processing various data types such as images, audios, text, 3D points, graphs, and events. (Image samples are taken from ImageNet dataset. Audio waveform visualizes a sample from TIDIGITS dataset. 3D points visualize a sample from ModelNet10 dataset. Graph sample is from CORA dataset.

5.1.2. Challenges

Major challenges can be categorized at different levels.

Device level: The ionic nature of resistive switching, although benefits data retention, imposes challenges to programming precision, energy and speed. The programming precision limits the representation capability of the resistive switch, or equivalently how many bits a device can encode. In addition, the programming energy and speed impact online learning performance. In addition, the degradation of the representation capability is further intensified by the read noise, manifestation by the current fluctuation under a constant voltage bias.

Circuit level: Analog resistive memory arrays are mostly interfaced with up- and downstream digital modules in a computing pipeline. As such, there is inevitable signal acquisition and conversion cost, which leads to the question of how to trade off between signal acquisition rate, precision, and power consumption. In addition, the parasitic resistance and capacitance, like the non-zero wire resistance, incur the so-called IR drop in the resistive memory crossbar array.

Algorithm level: So far many applications of resistive memory suffer from significant performance loss in the presence of resistive memory nonlinearities (e.g. noises), thus defeating their efficiency advantage over alternative digital hardware.

5.1.3. Potential Solutions

Device level: Various approaches are used to address the programming stochasticity, such as the local confinement of conducting filament ¹⁹⁶.  A denoising protocol using sub-threshold voltages has recently been developed to suppress the fluctuation of the device state and achieve up to 2048 conductance levels¹⁹⁷. In addition, homogenous switching may suppress stochasticity at the cost of larger program energy and time overheads⁶⁴. In terms of programming energy, small redox barriers and large ion mobilities may reduce switching energy and accelerate switching speed, at the expense of retention and thermal stability though.Circuit level: Typically, resistive in-memory computing relies on Ohm’s law and Kirchoff’s current law, resulting in current summation. However, there is a recent surge of interest in replacing current summation by voltage summation, which lowers down the static power consumption by eliminating current summation incurred Joule heating. In addition, fully analog neural networks have been proposed to get rid of the frequent analog-to-digital and digital-to-analog conversions ¹⁹⁸. To combat with the parasitic wire resistance, a simple solution is to increase device resistance in both ON and OFF states, such as that demonstrated in a 256×256 in-memory computing macros ¹⁹⁷.

Algorithm: A recent trend is hardware-software codesign to leverage resistive memory nonlinearities and turn them into advantages. For example, the programming stochasticity can be exploited by neural networks of random features (e.g. echo state networks ^165,166 and extreme learning machines¹⁹²) and Bayesian inference using Markov Chain Monte Carlo (MCMC) such as Metropolis–Hastings algorithm¹⁹⁹. Also, such programming noise is a natural regularization to suppress overfitting in online learning²⁰⁰. Moreover, hyperdimensional computing¹⁸⁹ and mixed-precision design such as high-precision iterative refinement algorithm paired with low-precision conjugate gradient²⁰¹ can withstand resistive memory programming noise. The reading noise can also be exploited for solving combinatorial optimization problems using simulated annealing, serving as a natural noise source to prevent the system from falling into the local minimum^202,203.

5.1.4. Conclusion

The fast advent of resistive switch-based in-memory computing in the last decade has demonstrated a wide spectrum of applications in machine learning and neuromorphic computing, reflected by its handling of different types of data.

However, there are still plenty of room, at device, circuit, and algorithm levels, to improve, which will help to fully unleash the power of in-memory computing with resistive switches and potentially yield a transformative impact on future computing.

5.2 Phase change materials

Abu Sebastian, Ghazi Sarwat Syed

5.2.1.Introduction

Phase-change memory (PCM) is arguably the most advanced memristive technology. Similar to conventional metal-oxide based memristive devices, information is stored in terms of changes in atomic configurations in a nanometric volume of material and the resulting change in resistance of the device ²⁰⁴. However, unlike the vast majority of memristive devices, PCM exhibits volumetric switching as opposed to filamentary switching. The volumetric switching is facilitated by certain material compositions along the GeTe-Sb₂Te₃ pseudo-binary tie line, such as Ge₂Sb₂Te₅ that can be switched reversibly between amorphous and crystalline phases of different electrical resistivity ²⁰⁵. Both transitions are Joule-heating assisted. The crystalline to amorphous phase transition relies on a melt-quench process whereas the reverse transition relies mostly on crystal growth.

Figure 10: Operational regimes of a phase change device when used for neuromorphic computing. On the right plot, the direct overwrite regime utilizing melt-quench dynamics is illustrated. The programming curves display the achievable conductance values in response to partial RESET pulses of varying amplitudes. As the RESET pulse is increased in amplitude, a larger amorphous volume is created mostly independent of the phase configuration prior to the application of the pulse. On the left plot, the characteristic accumulative property is demonstrated. It showcases the evolution of conductance values over successive applications of a SET pulse with a constant amplitude. As the amorphous region reduces in size due to crystallization dynamics, the device conductance progressively increases. Multiple experimental traces are overlaid in both plots.  

There are essentially two key properties that make PCM devices particularly well suited for neuromorphic computing (see Figure 10) ²⁰⁶. Interestingly this was pointed out by Stanford Ovshinsky, a pioneer of PCM technology, way back in 2003 when PCM was being considered just for memory applications ²⁰⁷. The first property is that PCM devices can store a range of conductance values by modulating the size of the amorphous region typically achieved by partial RESET pulses that melt and quench a certain volume of PCM material. This analogue storage capability, combined with a crossbar topology, allows for matrix-vector multiply (MVM) operations to be carried out in O(1) time complexity by leveraging Kirchhoff’s circuit laws. This makes it possible to realize an artificial neural network on crossbar arrays of PCM devices, with each synaptic layer of the DNN mapped to one or more of the crossbar arrays ^67,208. The second property referred to as accumulative property results from the progressive crystallization of PCM material upon application of an increasing number of partial SET pulses (see Figure 1). It is used for implementing DNN training ²⁰⁹, temporal correlation detection ²¹⁰, continual learning ²¹¹, local learning rules such as spike-timing-dependent plasticity ^212,213 and neuronal dynamics ²¹⁴.

PCM is at a very high maturity level of development and has been commercialized as both standalone memory²¹⁵ and embedded memory ²¹⁶. This fact, together with the ease of embedding PCM on logic platforms (embedded PCM) ²⁰⁸ make this technology of unique interest for neuromorphic computing.

5.2.2.Challenges

PCM devices offer write operations in the tens of nanosecond timescale which is sufficient for most neuromorphic applications in particular those targetting deep learning inference. The cycling endurance could also exceed a billion cycles (dependent on the device geometry) which is several orders of magnitudes higher than commercial Flash memory²¹⁷. This is sufficient for deep learning inference applications. The cycling endurance for partial SET pulses is much higher than that for full SET-RESET cycling and hence is widely considered sufficient for other neuromorphic applications as well. The read endurance is almost infinite for PCM when sufficiently low read bias is applied. Another key attribute is retention which is typically tuned through material choice ²¹⁸. However, the use of analogue conductance states in neuromorphic computing makes the retention time of intermediate phase configurations even more important which could be substantially lower than that of fully RESET states.

One of the primary challenges for PCM is integration density. For example, for DNN inference, it is desirable to have at least 10-100 Million on-chip weight capacity. The crossbar array for neuromorphic computing comprises metal lines intersected by synaptic elements, which are composed of one or more PCM devices and selector devices. Access devices such as bipolar junction transistors or metal-oxide-semiconductor field effect transistors are preferred for accurate programming, while two-terminal poly-silicon diodes offer scalability. To achieve high memory density, stacking multiple crossbar layers vertically is beneficial. BEOL selectors such as ovonic threshold switches show promise but face challenges in achieving precise current control. Edge effects and thermal cross-talk between neighboring cells become significant at smaller feature sizes ^219,220,221.

Figure 11: PCM non-idealities. (a) Device data after programming shows variability, reflected in broad distributions of analogue conductance values due to programming inaccuracies, read noise and drift variability (the top panel). Temperature increase raises state conductivity due to thermal carrier excitation and accelerates structural relaxation. (b) The conductance fluctuations manifest as synaptic weight noise, here shown as additive noise in terms of the percentage of the maximum synaptic weight. Using two PCM devices per synapse reduces this error. (c) The accumulative behaviour exhibit significant stochasticity mostly attributed to variations in the crystallization kinetics.

5.3 Ferroelectric materials

Thomas Mikolajick, Stefan Slesazeck and Beatriz Noheda

5.3.1.Status 

Ferroelectric materials are, in theory, ideally suited for information storage tasks since their switching is purely field-driven, holding the promise of extremely low write energy, and non-volatile at the same time. Moreover, unlike competing concepts like resistive switching or magnetic switching, ferroelectric materials offer three different readout possibilities giving a lot of flexibility in device design ²²². In detail, the following read schemes can be applied (see also middle part of Fig. 12):

• Direct sensing of the switched charge during polarization reversal, as used in the ferroelectric RAM (FeRAM) concept, results in a cell design similar to a dynamic random-access memory (DRAM) ²²³.

• Coupling of the ferroelectric to the gate of a field effect transistor and readout of the resulting drain current, as used in the ferroelectric field effect transistor (FeFET). This results in a cell that is similar to classical transistor-based charge storage (floating gate or a charge trapping) memory cell, which is most prominently used in Flash memories ²²⁴.

• Modulation of the tunnelling barrier in a ferroelectric tunneling junction (FTJ). As a result, we can realize a two-terminal device, which is essentially a special version of a resistive switching memory cell (see Chapter 5.1) ²²⁵.

Each of the mentioned read-out schemes has advantages and disadvantages and, therefore, the flexibility to use one of the three is a plus, especially in applications that go beyond pure memories, like neuromorphic computing.

However, traditionally ferroelectricity was only experienced in chemically complex materials, such as lead-zirconium titanate (PZT), strontium bismuth tantalate (SBT) or bismuth ferrite (BFO) which all are very difficult to incorporate into the processing flow for integrated electronic circuits, due to their limited stability in reducing environments. Another pervasive issue for the integration of ferroelectrics is their tendency to depolarize upon downscaling, an issue that is accentuated by their high permittivity. Organic ferroelectrics, the most prominent example being Polyvinylidenfluorid (PVDF), can mitigate this problem, as their low permittivities reduce the depolarization fields; while a rather high coercive field increases the stability of the polarization state. Such materials are ideally suited for lab scale demonstrations of new device concepts, due to their simple fabrication using a solution-based process, and are highly preferred for flexible and biocompatible electronics ²²⁶. However, their limited thermal stability has taken them out of the game for devices in integrated circuits. Therefore, although the technology in form of FeRAM ²²⁷ is on the market for more than 25 years, it has lacked the ability to scale in a similar manner as conventional memory elements and, therefore, it is still limited to niche applications that require a high re-write frequency together with nonvolatile like in data logging applications.

5.3.2.Challenges

With the discovery of ferroelectricity in hafnia (HfO₂) and zirconia (ZrO₂), the biggest obstacle of the limited compatibility with integrated circuit fabrication could be solved ²²⁸. HfO₂ and ZrO₂ are stable both in reducing ambient and in contact with silicon and their fabrication using established atomic layer deposition processes is standard in modern semiconductor process lines.  However, new difficulties, especially with respect to reliability ²²⁹, need to be solved. Challenges in this direction are aggravated by the metastable nature of the ferroelectric phase, which appears mostly at the nanoscale, making a full understanding of the polar phase quite demanding. While their high coercive field makes them very stable with respect to classical retention, the ferroelectric phase typically exists together with other non-polar phases, which prevents them from reaching the predicted polarization values (of the order of 50 uC/cm²) ²³⁰ . Moreover, the most serious problem of any nonvolatile ferroelectric device, the imprint, becomes very complex to manage in hafnia/zirconia-based ferroelectrics. Imprint is a shift of the hysteresis loops due to an internal bias. While this effect leads to a classical retention of the stored state that may look perfect, after switching, retention will be degraded and fixing the so-called opposite-state retention loss needs to be carefully done by material and interface engineering. Moreover, the high coercive field in this materials class becomes a problem as HfO₂ and ZrO₂ often show a pronounced wake-up and fatigue behavior and the field-cycling endurance is in many cases limited by the dielectric breakdown of the material.

While the issues mentioned so far are valid for any nonvolatile device application, in neuromorphic systems additional challenges arise, including the linearity of the switching behavior and tuning of the retention to achieve both short-term and long-term plasticity, as well as specific effects to mimic neurons, like accumulative switching ^{231 222}, which need to be explored using material and device design measures. Finally, large-scale neuromorphic systems will require a high integration density that demands 3-dimensional integration schemes, either realized by the punch-and-plug technology well-known from NAND Flash or by integrating devices into the back-end of the line.

5.3.3. Potential Solutions

Since the original report on ferroelectricity in hafnium oxide ²²⁸ , the boundary conditions for stabilizing the ferroelectric phase have been much better understood, although there are still a number of open questions. The goal is to achieve a high fraction of the ferroelectric phase without dead layers of non-ferroelectric phases at the interface to the electrodes or in the bulk of the film. This needs to be done under the boundary conditions of a realistic fabrication process, which means that sophisticated methods to control the crystal structure based on epitaxial growth are not possible. Epitaxial growth can help clarify scientific questions but the achieved results need to be transferred to chemical vapor deposition (CVD), including most prominently atomic layer deposition (ALD), or physical vapor deposition (PVD) processes using electrodes like TiN or TaN that can be integrated into electronic processes.

In the last years, it became obvious, that oxygen vacancies are, on the one hand, required to stabilize the ferroelectric phase  ²³² and, on the other hand,  detrimental to both the imprint and the field cycling behavior ²³³. Therefore, many proposals to integrate  the ferroelectric layer with additional thin layers in the film stack have been made and currently a lot of work is going in that direction. Moreover, it is clear, that the interface to the electrodes needs careful consideration. In this direction ²³⁴, facilitating the transport of oxygen, not only in the ferroelectric layer but also across the electrode interfaces, by minimizing the strain effects, may be the key to improving device performance   ²³⁵ . When it comes to structures that are in direct contact with silicon, a recent observation of a quasi-epitaxial growth of extremely thin hafnium-zirconium oxide films on silicon could be an interesting direction ²³⁶.  For concrete neuromorphic applications, the rich switching dynamics can be very helpful (see Fig. 12) ²³⁷. While in large devices a continuous switching between different polarization states is possible, devices scaled in the 10nm regime show abrupt and accumulative switching  ²³¹ . The former can be used for mimicking synaptic functions while the latter is helpful to mimic neurons. In classical nonvolatile memories, the depolarization fields created by non-ferroelectric layers or portions of the layer in series to the ferroelectric are a concern for the retention of the device. However, when creating short and long-term plasticity in synaptic devices, this can be turned into an advantage such that the device retention can be tailored.

Figure 12: Ferroelectric materials (center) enable three different basic memory cells (middle ring). These can be used in various ways in neuromorphic circuits (examples see outer part of the figure). The rich switching dynamics of ferroelectrics allow to tailor new devices mimicking neurons and synapses in a much more flexible and area efficient way as compared to their pure CMOS counterparts.

5.4 Spintronic materials for neuromorphic computing

Bernard Dieny, Tuo-Hung (Alex) Hou

5.4.1. Status

Spintronics is a merging of magnetism and electronics in which the spin of electrons is used to reveal new phenomena and implement them in devices with improved performances and/or new functionalities. Spintronics already found many applications for magnetic field sensors in particular in hard disk drives and more recently as non-volatile memory (MRAM) in replacement of e-FLASH and last-level CACHE memory. Spintronics can also bring very valuable solutions in the field of neuromorphic computing both as artificial synapses or neurons.

Artificial synapses are devices supposed to store the potential weight of the bounds linking two neurons. Various types of spintronic synapses have been proposed and demonstrated ⁶⁵. They are magnetoresistive non-volatile memory cells working either as binary memory, multilevel memory, or even in an analogue fashion. Their resistance depends on the history of the current that has flown through the device (memristor). Most of these devices are based on magnetic tunnel junctions (MTJ) which basically consist of two magnetic layers separated by a tunnel barrier. One of the magnetic layers has a fixed magnetization (the reference layer) whereas the magnetization of the other (the storage layer) can be changed by either pulse of magnetic field or current using phenomena such as spin transfer torque (STT) or spin-orbit torque (SOT)²³⁸. The resistance of the device depends on the amplitude and orientation of the magnetic moment of the storage layer relative to that of the reference layer (tunnel magnetoresistance effect: TMR). For a binary memory as in STT-MRAM or SOT-MRAM, only the parallel and antiparallel magnetic configurations are used ²³⁹. For multilevel or analogue memory, several options are possible as illustrated in  Fig.13. One consists of varying the proportion of the storage layer area which is in parallel or antiparallel magnetic alignment with the reference layer magnetization. This can be achieved by step-by-step propagating a domain wall within the storage layer using the STT produced by successive current pulses  (Fig.13a) ²⁴⁰, or by gradually switching the magnetization of the storage layer exchange coupled to an antiferromagnet using the SOT produced by the pulsed current flow in the antiferromagnet (Fig.13c)²⁴¹ or by gradually switching the grains of a granular storage media similar to the ones used in hard disk drives  (Fig.13d) ²⁴², or by nucleating a controlled number of magnetic spin nanotextures in the storage layer such as skyrmions (Fig.13e) ²⁴³. Alternatively, the memristor resistance can also be varied by changing the relative angle between the magnetization of the reference and storage layers using all intermediate angles between 0° and 180° instead of only parallel and antiparallel configurations  (Fig.13b)  ¹⁰⁴. Chains of binary magnetic tunnel junctions can also be used to achieve spintronic memristors but at the expense of larger footprint ²⁴⁴.

Figure 13: Various realizations of spintronic memristors: (a) based on domain wall propagation in the storage layer; (b) based on variation of angle between storage layer and reference layer magnetization; (c) based on SOT in ferromagnetic storage layer exchange coupled to an antiferromagnetic SOT line; (d) implementing a storage layer made of a granular layer similar to the one used in recording technology; (e) based on a controlled number of skyrmions nucleated in the storage layer.

Concerning artificial neurons, the conventional CMOS neuron circuit is limited by its large area because a large number of transistors and a large-area membrane capacitor is required for implementing Integrate-and-Fire I&F functions ²⁴⁵. Recently, several spintronic neuron devices were reported to generate spike signals by leveraging nonlinear and stochastic magnetic dynamics without the need for additional capacitors and complex peripheral circuitry. Spintronic neurons potentially show a great advantage for compact neuron implementation ²⁴⁶.

Assembly of interacting spin-torque nano-oscillators (STNO) based on the structure of magnetic tunnel junctions (MTJs) was proposed to achieve neuron functionality. Unstable conductance oscillation that mimics spike generation is induced at hundreds of MHz to several tens of GHz by flowing a current through the device. The frequency and amplitude of oscillation vary with the applied current and magnetic field. Torrejon et al. demonstrated spoken-digit or vowel recognition using such an array of nanoscale oscillators ²⁴⁷.

Superparamagnetic tunnel junctions can also be used to mimic stochastic neurons. They have much lower thermal stability compared to the MTJ used for memory so they stochastically switch between antiparallel (AP) and parallel (P) states due to thermal fluctuations, which is referred to as telegraphic switching ²⁴⁸. This switching mode can be used to generate Poisson spike trains in spiking neural networks (SNNs) as well as for probabilistic computing ²⁴⁹. The MTJ device with high thermal stability, which can implement not only synapses but also neurons in an all-spin neural network, was proposed by Wu et al. ²⁵⁰. The reduction of the thermal stability factor is induced by self-heating at high bias voltage for neuron operations ²⁵¹. At low bias, it stably stores weight information as synapses.

Many other new spintronic materials and mechanisms were also investigated for the feasibility of neuron devices, in particular based on magneto-electric effects. For instance, by playing with magneto-ionic effects influencing the anisotropy at magnetic metal/oxide interfaces, the density of skyrmions ²⁵² and even their chirality could be controled electrically ²⁵³.  Jaiswal et al. designed a magnetoelectric neuron device for SNNs  ²⁵⁴ . Zahedinejad et al demonstrated that electrically manipulated spintronic memristors can be used to control the synchronization of spin Hall nano-oscillators for neuromorphic computing ²⁵⁵.

5.4.2. Challenges

Building useful fully functional neuronal circuits requires large-scale integration of layers of artificial neurons interconnected with spintronic synapses. Crossbar architectures can achieve cumulate and multiply functions very efficiently in an analogue manner. An advantage of magnetic tunnel junctions over other technologies based on materials such as resistive oxide or phase change is their write endurance associated with the fact that their resistance change does not involve ionic migration. However, they exhibit a lower R_OFF/R_ON ratio (~4 for MRAM versus 10 to 100 for RRAM or PCRAM) but also narrower cell-to-cell distribution of resistance in R_OFF and R_ON states. In crossbar architecture, MTJ should have high resistance to minimize power consumption. Therefore efforts should be pursued to further increase the TMR amplitude of MgO-based MTJ and bring it closer to the expected theoretical values of several 1000% ²⁵⁶. In high-resistance MTJ, other approaches such as SOT or voltage control of anisotropy (VCMA) could be used to change the MTJ resistance. In all cases, the control of the resistance change induced by current or voltage pulses must be improved. The operating temperature has often also a significant impact on the magnetic properties which imposes challenges on system design.

Concerning artificial neurons, the DC power required to trigger the magnetization dynamics of STNO neurons is still relatively high (mW range) ²⁵⁷. Ways must be found to reduce it by using different materials or new designs. The switching speed and endurance in superparamagnetic tunnel junctions and self-heating-assisted MTJ neurons could be further enhanced to improve processing speed and system reliability ²⁵⁸²⁵⁹. How to continue improving variability across millions of synapses and thousands of neurons to ensure high accuracy in future neuromorphic systems remains an actively research topic. Interconnecting all these devices is also a challenge and innovative approaches beyond classical interconnects must be found notably by taking advantage of 3D integration.

5.4.3. Potential Solutions

STT-MRAM entered volume production in 2019 at major microelectronic companies²⁶⁰. This marked the adoption of this hybrid CMOS/magnetic technology by the microelectronic industry. Thanks to the combined efforts of the chip industry, equipment suppliers, and academic laboratories, spintronics is progressing very fast. Material research is very important to increase magnetoresistance amplitude, switching currents, STT and SOT efficiency, VCMA efficiency, reduce dependence on operating temperature, reduce current to trigger oscillations in STNO, and reduce disturbance due to parasitic field. Investigations are in progress involving antiferromagnetic materials for reduced sensitivity to the field and access to THz frequency operation ²⁶¹ , half-metallic materials such as Heusler alloys for enhanced TMR amplitude and reduced write current ¹⁰⁵ , topological insulators for very efficient spin/charge current interconversion possibly combined with ferroelectric materials ²⁶² .

Concerning interconnects, fortunately, magnetic materials are grown in back-end technology and can be stacked but at the expense of complexity and cost. Long-range information transmission can be carried out via spin-current or magnons or by propagating magnetic textures such as domain walls ²⁶³ or skyrmions ²⁶⁴. Light could also be used to transmit information in conjunction with recent developments related to all-optical switching of magnetization ²⁶⁵. Besides, a great advantage of spintronic stacks is that they can be grown on almost any kind of substrates provided the roughness of the substrate is low enough compared to the thickness of the layers comprised in the stack. This enables the use of the third dimension by stacking several spintronic structures thereby gaining in interconnectivity ²⁶⁶.

5.4.4. Concluding remarks

Spintronics can offer valuable solutions for neuromorphic computing. Considering that STT-MRAM is already in commercial production, it is very likely that the first generation of spintronic neuromorphic circuits will integrate this technology. Next, crossbar arrays implementing analogue MTJ may be developed as well as neuronal circuits based on the dynamic properties of interacting STNO for learning and inference.   Still, many challenges are on the way towards practical applications, including speed, reliability, scalability, and variation tolerance which need to be addressed in future research.

5.5 Optoelectronic and photonic implementations

Akhil Varri, Frank Brückerhoff-Plückelmann, Wolfram Pernice

5.5.1. Status

Computing using light offers significant advantages in highly parallel operation exploiting concepts like wavelength and time multiplexing. Moreover, optical data transfer enables low power consumption, better interconnectivity, and ultra-low latency. Already in the 1980s, first prototypes were developed, however, the bulky tabletop experiments could not keep pace with the flourishing CMOS industry. Nowadays, novel fabrication processes and materials enable the (mass) production of photonic integrated circuits, allowing photonic systems to compete with their electronic counterparts. Especially in the area of data-heavy neuromorphic computing, the key advantages of photonic computing can be exploited.

Scientific efforts in neuromorphic photonic computing can be segregated in two major directions: (i) one approach is building hardware accelerators that excel at specific tasks, e.g., computing matrix-vector multiplications, by partially mimicking the working principles of the human brain; and (ii) designs which aim to emulate the functionality of biological neural networks. Such devices are able to replicate the behavior of a neuron, synapse, learning mechanisms and ultimately implement a spiking neural network.

There has been considerable progress in the (i) direction leading back to 2017 when Shen et al. ²⁶⁷ demonstrated vowel recognition where every node of the artificial neural network is physically represented in the hardware using a cascaded array of interferometers. This scheme has also been scaled to implement a 3-layer deep neural network with in-situ training capability ²⁶⁸. In addition, Feldmann et al. ²⁶⁹ have demonstrated neurosynaptic networks on-chip and used them to perform image recognition. The photonic circuit deploys non-volatile phase change material (PCM) to emulate the synapses and exploits the switching dynamics as a non-linear activation function. As highlighted in section 5.2, integration of PCMs also lead to in-memory computing functionality owing to their nonvolatile nature.

For the (ii) direction, significant work has been done on a device level to mimic individual components of the brain. Excitable lasers combining different material platforms such as III-V compounds, and graphene have been shown to demonstrate leaky integrate and fire-type characteristics of a neuron ⁷⁴. Also, neurons based on optoelectronic modulators have been shown in the literature. For synapses, photonic devices combined with PCMs, amorphous oxide semiconductors, and 2D materials have been used to demonstrate synaptic behavior such as spike-time-dependent plasticity, memory, etc ^74,270,271.

In the following, we break down the challenge of building neuromorphic photonic hardware to various subtopics, ranging from increasing the fabrication tolerance of the photonic circuit to co-packaging the optics and electronics. Then, we review the current advances in those areas and provide an outlook on the future development of neuromorphic photonic hardware.

5.5.2. Challenges

Figure 14: An illustration of an opto-electronic system capable of performing computing operations. The electro-optic modulator (EOM) encodes the input data from the memory or a real-time sensor into the light fields. The encoded light fields are then coupled to a photonic processor which emulates the neurosynaptic behavior. The results from the processor is passed on to detectors which transform the signals back into the electronic domain and are finally stored in a memory.  

A major challenge is combining the various building blocks shown in Figure 14. Silicon on insulator is the platform of choice for building large circuits owing to the matured CMOS process flow and the high refractive index contrast between the silicon waveguide and oxide cladding. However, silicon has no second-order nonlinearity as it is centrosymmetric. Further, silicon being an indirect bandgap material cannot emit light. This strongly limits the options for implementing nonlinear functions and spiking dynamics crucial for an all-optical neural network. Therefore, most of the research on mimicking neurons is focused on novel material platforms that support gain. A key challenge is integrating those different material platforms. For example, a circuit may deploy neurons based on III-V semiconductor heterojunction and synapses that are built with PCMs on silicon. Therefore compact and fabrication error tolerant optical interconnects are crucial for the performance of the whole system.

Apart from packaging various optical components, the electro-optic interface imposes an additional challenge. Typically the input data is provided by digital electrical signals whereas optical data processing is analog. This requires analog-to-digital converters (ADCs) and digital-to-analog converters (DACs) for digital systems to interface with the chip as shown in Figure 14. For large circuits, co-packaging electronics and photonics increases the footprint and cost significantly. This negatively affects the throughput.

In addition to the above, fabrication imperfections will also impact the performance of a photonic circuit. Components such as ring resonators, cavities, interferometers, etc. employed in many photonic circuits are designed to operate at a certain wavelength. But, due to factors such as etching rate, sidewall angle, and surface roughness, the wavelength of operation many times does not match with the design. Hence, in many cases, active methods such as thermo-optic phase shifters are employed to post-fabricate trim the wavelength. This results in unnecessarily increased electronic circuitry adversely affecting the scalability of the system.   

Lastly, a challenge that may be critical in the future is the electro-optic modulator (EOM) efficiencies that depend on the material properties and configuration. An important figure of merit for EOM efficiency is VpiL. This merit shows the voltage that needs to be applied and the length of the modulator required to obtain a pi phase shift to the input. A smaller merit figure suggests increased power efficiency and a compact footprint. As photonic neuromorphic circuits are supposed to scale up in the future, the power budget and space available on-chip will play an essential role in influencing the designs.

5.5.3. Potential Solutions

Solutions addressing the challenges mentioned above lie on multiple fronts. First, we discuss how the scalability can be improved from a device-level perspective. The compact footprint, power efficiency, and cascadability of the neurons are essential characteristics for improving the scalability. In this regard, modulator-based neurons can be improved by integrating with materials such as electro-optic polymers. These materials have an order of magnitude higher r33 electro-optic coefficient compared to bulk lithium niobate which has been conventionally the material of choice for modulators. As a result, electro-optic polymers integrated with silicon waveguides show very low VpiL among fast modulators ²⁷². In addition, novel materials such as epsilon-near-zero (ENZ) which are promising for optical nonlinearity can also be explored ²⁷³. Nevertheless, for the widespread use of these devices, a better understanding of the material properties and engineering efforts to integrate them into the existing manufacturing process flow is required.

Particularly, integration techniques such as micro-transfer printing, flip-chip bonding, and photonic wire bonding will play a key role. To solve the problem of packaging with electronics, strategies such as monolithic fabrication where the photonics and electronics are on the same die need to be investigated. Foundries are now offering multi-project wafer runs with these state-of-the-art packaging techniques.

For improving the scalability of spike-based processing systems, another class of neurons that is very promising is the vertical cavity surface emitting lasers (VCSELs). VCSELs can integrate 100 picosecond-long pulses and fire an excitable spike when the sum crosses a certain threshold emulating biological neurons. Recently, it was shown that the output of one layer of VCSEL neurons combined with a software-implemented spiking neural network can perform 4-bit image recognition ²⁷⁴. In order to build the entire system on hardware and perform larger experiments, 2D VCSEL arrays flip-chip bonded on a silicon die can be examined.  

Finally, to address the challenge of fabrication imperfections, passive tuning approaches can be of interest which need no additional circuitry and is non-volatile. One direction could be the use of phase change materials such as GaS, and Sb2Se3 to correct for the variability in photonic circuits ²⁷⁵. These materials are very interesting since their real part of the refractive index can be tuned while keeping low absorption at telecom wavelengths. Another approach for post-fabrication passive trimming could be to use an electron beam or ion beam to change the material properties of the waveguide. This method is also scalable as these tools are widely used in the semiconductor industry.

5.5.4. Concluding remarks

Applications like neuromorphic computing are particularly promising for optics where its unique advantages (i.e high throughput, low latency, and high power efficiency) can be utilized. Presently, there have been instances in literature where different devices have been proposed to emulate the individual characteristics of a neurosynaptic model. However, there is a lot of scope for research in materials science to pave the way toward more compact, cascadable, and fabrication-friendly implementations. Further, large-scale networks are expected to scale in the near future by integrating state-of-the-art packaging techniques that are now available to research groups and startups.

To summarize, the growth of integrated photonics has led to a resurgence of optical computing not only as a research direction but also commercially. It is exciting to see how the field of neuromorphic photonics will shape as advancements in science and technology continue to happen.    

5.6 2D materials

Mario Lanza, Xixiang Zhang, Sebastian Pazos

5.6.1. Status

Multiple studies have claimed the observation of resistive switching (RS) in two-dimensional layered materials (2D-LMs), but very few of them reported excellent performance (i.e., high endurance and retention plus low switching energy, time and voltages) in a reliable and trustable manner, and in a device small enough to be attractive for high-integration-density applications (e.g., memory, computation).

The best RS performance observed in 2D-LMs is based on out-of-plane ionic movements. In such types of devices, the presence and quality of the RS phenomenon mainly depend on three factors: the density of native defects, the type of electrode used, and the volume of the dielectric (thickness and area). In general, 2D-LMs with excellent crystallographic structure (i.e., without native defects, such as those produced by mechanical exfoliation) do not exhibit stable resistive switching. Reference ²⁷⁶ reported that mechanically exfoliated multilayer MoS₂ does not show RS; only after oxidizing it (i.e., introducing defects) it shows RS based on the migration of oxygen ions. Along these lines, reference ²⁷⁷ showed that mechanically exfoliated multilayer hexagonal boron nitride (h-BN) does not exhibit RS; instead, the application of voltage produces a violent dielectric breakdown (DB) followed by material removal. The more violent DB phenomenon in h-BN compared to MoS₂ is related to the higher energy for intrinsic vacancies formation: >10 eV for boron vacancies in h-BN versus <3 eV for sulfur vacancies in MoS₂. Some articles claimed RS in mechanically exfoliated 2D-LMs, but very few cycles and poor performance were demonstrated; those observations are more typical of unstable DB than stable RS. Reference ²⁷⁸ reported good RS in a crossbar array of Au/h-BN/graphene/h-BN/Ag cells produced by mechanical exfoliation, but in that study, the graphene film shows amorphous structure in the cross-sectional transmission electron microscope images. Hence, stable and high-quality RS based on ionic movement has never been demonstrated in as-prepared mechanically exfoliated 2D-LMs. This is something expected because ionic-movement-based RS is only observed in materials with high density of defects (e.g., high-k materials, sputtered SiO₂), but not in materials with low density of defects (e.g., thermal SiO₂), as the higher energy-to-breakdown forms an irreversible DB event.

On the contrary, 2D-LMs prepared by chemical vapor deposition (CVD) and liquid phase exfoliation (LPE) have exhibited stable RS in two-terminal memristors ²⁷⁹ and three-terminal (memtransistors) configurations ²⁸⁰, though in the latter the switching mechanism is largely different. In two-terminal devices . In such cases, the RS is enabled by the migration of ions across the 2D-LM. In transition metal dichalcogenides (TMDs), the movement of chalcogenide ions can be enough to leave behind a metallic path (often referred to as conductive nanofilament or CNF) that produces the switching (similar to oxygen movement in metal-oxides) ²⁷⁶. However, in h-BN metal penetration from the adjacent electrodes is needed, as this material contains no metallic atoms ²⁷⁷. In 2D-LMs prepared by CVD and LPE methods ionic movement takes place at lower energies (than in mechanically exfoliated ones) due to the presence of native defects (mainly lattice distortions and impurities). The best performance so far has been observed in CVD-grown ~6-nm-thick h-BN, as it is the only material enough insulating and thick to keep low the current in the high resistive state ⁷³. This includes the coexistence of bipolar and threshold regimes (the second one with highly-controllable potentiation and relaxation), bipolar RS with endurances >5 million cycles (similar to commercial RRAM memories and phase-change memories) ²⁸¹, and ultra-low switching energies of ~8.8 zJ in threshold regime ²⁸². Moreover, high yield (~98%) and low variability has been demonstrated ²⁸². In 2D-LMs produced by LPE or other solution-processing methods ²⁸³, the junctions between the flakes and their size play a very important role, and while there is evidence of potentially good endurance, synaptic behaviour and variability, sub-µm downscaling still hasn’t shown equivalent performance ²⁸⁴.

Apart from ionic movement, 2D-LMs can also exhibit RS based on ferroelectric effect ²⁸⁵. A remarkable example is -In₂Se₃, which has electrically switchable out-of-plane and in-plane electric dipoles. Recent works have demonstrated that RS in ferroelectric -In₂Se is ensured by three independent variables (polarization, initial Schottky barrier and barrier change), and that it delivers multidirectional switching and photon storage ²⁸⁶. However, the endurance and retention time are still limited to hundreds of cycles, and stable ferroelectric RS at the single-layer limit remain unexplored.  Finally, tunable optoelectronic properties and unique electronic structure attainable through 2D-LM heterostructures present enormous potential for near-/ in-sensor computing in neuromorphic systems. The responsiveness to physical variables (light, humidity, temperature, pressure, torsion) of 2D-LM memristor and memtransistor devices allows to mimic biological neurosynaptic cells (visual cortex, tactile receptors) ²⁸⁷.

5.6.2. Challenges and Potential Solutions

The main challenge of RS devices (of any type) is to exhibit high endurance in small devices. Many studies have reported RS in large devices with sizes >10 µm², and claimed that their devices are “promising” for memory and computing applications. This is a huge and unreasonable exaggeration; these two applications require high integration density, as commercial devices for those applications have sizes down to tens or hundreds of nanometers. It should be noted that in ionic-movement-based RS devices the CNF always forms at the weakest location of the sample; when the device size is reduced the density and size of defects is (statistically) reduced, which produces an increase of the forming voltage ²⁸⁸. Hence, the CNF of smaller devices is wider due to the larger amount of energy delivered during the forming. This has a huge effect on state resistances, switching voltages, time and energy, as well as endurance, retention time and device-to-device variability. In other words, the fact that a large device (>10 µm²) exhibits good RS does not mean that a small device (<1 µm²) made with the same materials will also exhibit it; hence, RS “promising for memory and computation” is only the one that is observed in devices with sizes of tens/hundreds of square nanometers.

Taking this into account, the main challenge in 2D-LMs based devices is to observe RS in small devices, and the most difficult figure-of-merit to obtain is (by far) the endurance. Reference ²⁸⁹ demonstrated good RS in 5 µm × 5 µm Au/h-BN/Au devices, in which the h-BN was ~6-nm-thick and grown by CVD method; however, when the size of the devices was reduced to 320 nm × 420 nm the yield and the number of devices observed was very limited. The main issue was the current overshoot during the switching, which takes place randomly and produced irreversible DB in most devices. Similarly, solution-processed Pt/MoS₂/Ti devices ²⁹⁰ showed excellent performance across all figures-of-merit observed in 25 µm² devices, but such performance has not been reported for 500 nm × 500 nm devices patterned via electron beam lithography. In this case, the large size of the nanoflakes (slightly below 1 µm minimum) may be imposing an intrinsic scaling limitation. Meanwhile, the scaling and overshoot problem was problem was solved in reference ⁷³ integrating CVD-grown h-BN right on top (via wet-transfer method) of a silicon complementary metal-oxide-semiconductor (CMOS) transistor, which acted as instantaneous current limitation. Moreover, this approach brings the advantage of a very small device size (in reference ⁷³ it was only 0.053 µm², as the bottom electrode of the RS device is the via from one of the metallization levels). The heterogeneous integration of the 2D-LMs at the back-end-of-line (BEOL) wiring of silicon microchips could be a good way of testing materials for RS applications and directly integrating selector devices with each memristor (into one transistor-one memristor 1T1M cells), which is fundamental for the realization of large memeristive synapse arrays — all state-of-the-art demonstrations of memristive neural accelerators based on mature memristor devices use 1T1M cells or differential implementations of such (2T2M, 4T4M). So far these CMOS testing vehicles for RS materials are mainly employed by the industry; in the future, wide spreading this type of testing vehicles among academics working in the field of RS could improve the quality of the knowledge generated.  In addition, these devices may benefit from common practices in the field of silicon microchip manufacturing, such as surface planarization, plug deposition and high-quality, thick interconnect techniques.

Next steps in the field of 2D-LMs for RS applications consist on improving the materials quality to achieve better reproducibility of the experiments (from one batch to another) and adjust the thickness and density of defects to achieve better figures-of-merit in nanosized RS devices while growing 2D-LM at the wafer-scale ²⁹¹. Recent studies successfully synthesized large-area single-crystal 2D-LMs via CVD ²⁹², although in most cases it is only monolayer. However, monolayer 2D-LMs are less than 1-nm-thick, and when they are exposed to an out-of-plane electrical field very high leakage current is generated even if no defects are present, which increases a lot the current in HRS and the energy consumption of the device. Reference ²⁹³ recently presented the synthesis of single-crystal multilayer h-BN using scalable methods, but controlling the number of layers is still difficult. Electrical studies in such types of single-crystal multilayer samples should be conducted. Improving manipulating methods to prevent the formation of cracks during transfer is also necessary, although it is worth mentioning that multilayer h-BN materials are more mechanically stable than monolayers.

Recent demonstration of vector-matrix multiplication using MoS₂ memtransistors ²⁹⁴ is a promising advance in terms of a higher-level functional demonstration, though the fundamental phenomenon exploited is the well-known floating gate memory effect, not unique to 2D-LM themselves. Meanwhile, understanding the role of flake size in the functionality of solution-processed 2D-LM two-terminal synaptic devices is critical to address the true scaling limitations of such approach, key aspect to define potential realistic applications in neuromorphic systems. On the other hand, sensing capabilities emerge with great potential for biological synaptic mimicking. The full potential of different 2D-LM material heterostructures and memtransistors opens a huge design-space worth of exploration. In that sense, the complex physical characteristics offered by different 2D-LM hold the potential not only for basic neuromorphic functionality but also for higher-order complexity. This could be exploited to achieve high-complexity neural and synaptic functions ¹⁰¹, more closely mimicking actual biological systems. However, in parallel to elucidating the physical properties and capabilities of these material systems, efforts should be put into strengthening the quality of the reported results, focusing on proper characterization methods, reliable practices, and statistical validation.

5.6.3. Concluding remarks

Leading companies like TSMC, Samsung, IBM and Imec have started to work with 2D-LMs, but mainly for sensors and transistors. In the field of 2D-LMs based neuromorphic devices most work is being carried out by academics. In this regard, unfortunately, many studies make a simple proof-of-concept using a novel nanomaterial without measuring essential figures-of-merit like endurance, retention and switching time. What is even worse, in many cases the studies employ unsuitable characterization protocols that heavily overestimate the performance (the most popular case is the erroneous measurement of endurance ²⁹⁵), withholding information regarding the failure mechanisms that lead to certain performance metrics not being achieved on some devices . This working style often result in articles with striking numbers (i.e., performance), but those are unreliable, and it is really bad for the field because it creates a hype of expectations and disillusion among investors and companies. The most important is that the scientists working in this field follow a few considerations: i) always aim to show high performance in small (<1 µm²) devices fabricated using scalable methods (even better if they are integrated on a functional CMOS microchip, not on an unfunctional SiO₂ substrate), ii) measure all the figures-of-merit of several (>100) memristive devices for the targeted application (this may vary depending on the application) ²⁹⁶, iii) define clearly the yield-pass criteria and the yield achieved, as well as the device-to-device variability observed, and iv) whenever a failure mode is observed preventing of reaching a desired figure-of-merit, clearly convey it to maximize the probabilities of finding a solution .

6. Materials challenges and perspectives

Stefan Wiefels, Regina Dittmann

6.1 Materials challenges

For the neuromorphic computing approaches addressed in chapter 3, the use of emerging memories based on novel materials will be key in order to improve their performance and energy-efficiency.   This chapter discusses the most relevant properties and challenges for different use cases and how they relate to the respective materials properties. However, it is important to note that a dedicated co-development of material with the read- out and write algorithm and circuitry will be required  in order to advance the field.

Figure 15: Materials challenges for neuromorphic computing. Novel NVMs need to be scalable, fast, reliable and allow for analog operation.

6.1.1.  Scaling 

One main driving force to use emerging materials and devices is to gain space and energy efficiency by the fabrication of highly dense cross-bar arrays. For STT-MRAM, scaling down to 11 nm cells has been demonstrated as well as the realization of 2 Mb embedded MRAM in 14 nm FinFET CMOS ²⁹⁷. However, due to the small resistance ratio of 2-3, the read-out of magnetic tunnel junctions (MTJs) is more complex than for other technologies. Nevertheless, a 64 x 64 MTJ array, integrated into 28 nm CMOS, has recently been realised ⁷⁶. Advancement from the material side will be needed in order to increase the resistance ratio of MTJs in the future.

For ferroelectric HfO₂-based devices, the main challenge with respect to scaling is to decrease the thickness reliably in order to enable 3D capacitors with 10 nm node and to obtain a uniform polarisation at the nanoscale of a material that currently still contains a mixture of different phases. Therefore, ultrathin films with the pure ferroelectric orthorhombic phase and without any dead-layers at the interfaces will be key to approach the sub-20 nm regime of hafnia-based ferroelectric devices ⁹.

PCM devices can be fabricated on the sub-10 nm scale ²⁹⁸. The limiting factor for CMOS integrated PCM devices is the high RESET current which is required to implement larger access transistors ²⁹⁹. Commercially available ReRAM cells with conventional geometries have been co-integrated on 28 nm CMOS technology. By employing  a sidewall technique and nanofin Pt electrodes, small arrays with 1 nm × 3 nm HfO₂ ²⁹⁹ cells  and 3 × 3 arrays of Pt/HfO₂/TiO_x/Pt cells  with a 2 nm feature size and a 6 nm half-pitch  have been fabricated, respectively ³⁰⁰.

With respect to ultimate scaling, the loss of oxygen to the environment might pose limitations to the retention times for ReRAM devices scaled in the sub-10 nm regime ³⁰¹. However, filaments in the size of 1-2 nm can be stable if they are stabilized by structural defects such as grain boundaries or dislocations. Therefore, finding materials solution for confining oxygen vacancies to the nanoscale might retain the required retention for devices in the few nm scale.

6.1.2. Speed 

Although the extensive parallelism leads to high demands for scaling, it is considered an advantage as it makes the race for ever increasing clock frequencies obsolete ³⁰². In contrast, the operation speed is closely linked to the respective application, i.e. the timing is based on real physical time ³⁰². As signals processed by humans are typically on a time scale of milliseconds or longer, the expected speed benchmark is well below the reported speed limits of emerging NVMs. Nevertheless, it is reasonable to understand the ultimate speed limits of NVM concepts in order to estimate maximum learning rates, to explore the impact of short spiking stimulation. Furthermore, novel computing concepts, as discussed in chapter 7, might still benefit from higher clock frequencies. For MRAM, reliable 250 ps switching has been demonstrated by using double spin-torque MTJs which consist of two reference layers, a tunnel barrier, and a non-magnetic spacer  ³⁰³. FeRAM arrays have successfully been switched with 14 ns at 2.5 V. Ferroelectric field effect transistors (FeFETs) have been shown to switch with < 50 ns pulses in 1 Mbit memory arrays ⁹. PCM devices can be switched with pulses < 10 ns ²⁹⁸. In general, their speed is limited by the crystallization time of the material. It has been shown exemplarily on Ge_xSn_yTe samples that this time can be tuned in a broad range of 25 ns up to 10 ms by adjusting the material composition ³⁰⁴. Thus, it has a high potential to match the operation time of an NC system to the respective application. For VCM ReRAM SET and RESET switching with 50 ps and 400 ps has been demonstrated ³⁰⁵. Both are so far limited by extrinsic effects and device failure modes rather than by intrinsic physical rate limiting steps.

6.1.3. Reliability 

Independent of the application, the reliability of the memory technology has to be taken into account. In the case of implementing NVMs as artificial synapses, the requirements of learning and inference phase have to be distinguished. Whereas the endurance is more relevant for learning schemes, the stability of the programmed state, i.e. the retention and robustness against read disturb have to be sufficient for reliable inference operations.

6.1.4. Endurance 

While MRAM has in principle unlimited endurance, all memristive devices that are based on the motion or displacement of atoms such as ReRAM, PCM and ferroelectric systems have limited endurance. For silicon-based FeFETs the endurance is typically in the order of 10⁵, which is mainly limited by a dielectric breakdown in the SiO₂ at the Si-HfO₂ interface ⁹. Regarding the endurance of VCM ReRAM, it has been demonstrated with convincing statistics that > 10^6 cycles are realistic. Some reports suggest maximum cycle numbers of more than 10^10 ²⁹⁵. Depending on the material system, various failure mechanisms for the endurance are discussed. The microstructure of the switching material might degrade or be irreversibly penetrated by metallic atoms ⁹. In VCM ReRAM an excessive generation of oxygen vacancies was discussed as endurance limiting factor ³⁰⁶. Novel material solutions which confine ions to the intended radius of action might be a pathway to increase the endurance of ReRAM devices. For PCM, it was suggested to implement multiPCM synapses. Arbitration over multiple memory elements might circumvent endurance as well as variability issues ⁹.

Typical limitations with respect to a reliable operation of ferroelectric memories is the so called wake-up effect which causes an increasing polarization after a few cycles and the fatigue resulting in a decrease of the polarisation for high cycle numbers. Both are induced by the motion of defects such as oxygen vacancies and will have to be tackled in the future by intense material research in this field.

6.1.5. Retention 

After training, the state of the non-volatile memory synapse is required to be stable for 10 years at an operating temperature of 85 °C. However, for many applications in the field of neuromorphic applications the requirement is much more relaxed in particular for the training phase. From a thermodynamical point of view, the states in ferroelectric or ferromagnetic memories might both be stable. In contrast, ReRAM and PCM devices store information in the configuration of atoms where both LRS and HRS are metastable states and the retention is determined by material parameters such as the diffusion coefficient of the respective species ⁹. Here, the degradation is not a digital flipping of states but a gradual process. For PCM, the drift of the resistance state is caused by the structural relaxation of the melt-quenched amorphous phase ²⁹⁹. Apart from a drifting of the state, a broadening of the programmed state distribution (e.g. resistance) is typically observed for ReRAM ³⁰⁷.  Further, since analogue or multi-level programming are highly relevant for NC, it should be considered that intermediate resistance states might have a reduced retention compared to the edge cases of high and low resistive states as demonstrated for PCM devices ⁹.

6.1.6. Read disturb 

During inference, frequent reading of the memory elements is required which should not change the learned state. For a bipolar ReRAM memory, a read disturb in the HRS/LRS occurs mainly when reading with a SET/RESET polarity since the read-disturb can be considered as an extrapolation of the SET/RESET kinetics to lower voltages. Nevertheless, the HRS state in bipolar filamentary VCM has been demonstrated by extrapolation to be stable for years at read voltages up to 350 mV ³⁰⁸.

6.1.7. Variability 

Variability is most pronounced for systems which rely on the stochastic motion and redistribution of atoms such as ReRAM and PCM.  Here, the variability from device to device (D2D), from cycle to cycle (C2C) as well as even from one read to the next (R2R) have to be distinguished. By optimizing fabrication processes, the D2D variability can be kept comparatively low. In contrast, the C2C variability for filamentary resistive ReAM and PCM can be significant due to the randomness of filament ³⁰⁸ or crystal ²⁹⁹ growth, respectively. However, using smart programming algorithms, the C2C variability can be very well reduced to a minimum  ³⁰⁸. But, R2R variations remain in form of read noise in filamentary VCM. It is typically attributed to the activation and deactivation of traps or the random redistribution of defects  ³⁰⁸ and strongly depends on the material ³⁰⁹. For PCM, R2R variations are caused by 1/f noise and temperature induced resistance variations. One approach to address these issues as well as the drift is to use the so called projected phase change memory with a non-insulating projection segment in parallel to the PCM segment.

Although the variability is a challenge for storage applications, it might be possible to design NC systems to exploit it ³⁰². In the end, a thorough understanding of the intrinsic variability might enable to match NC application and material ³⁰⁹ .

6.1.8 Analog operation 

For most computing concepts described in Chapter 3, their operation with binary memory devices is strongly limited and the possibility to adjust multiple states is of crucial importance. For devices with thermodynamically stable states such as ferroelectric or magnetic memory, intermediate states rely on the presence of domains. As a result, the performance strongly depends on the specific domain structure and scaling might be limited by the size of the domains. Nevertheless, multilevel switching has been demonstrated by fine tuning of programming voltages for both FTJ and FeRAM ³¹⁰.

For ReRAM and PCM, the metastable intermediate states have to be programmed in a reliable manner. Since these states are kinetically stabilized during programming, the success depends strongly on the switching kinetics of the specific system, the operation regime and the intrinsic R2R variability of the material. For PCM devices, intermediate states can be addressed by partial reset pulses which result in partial amorphisation. As a result of the crystallization kinetics, a gradual crystallization can be obtained by consecutive pulses.

Filamentary ReRAM devices usually undergo an abrupt SET which is caused by the self-accelerating, thermally-driven filament formation.  However, in a limited operation regime intermediate states can be obtained with a good control of the SET current, fast pulses or if the kinetics are slowed down. This is the case for non-filamentary systems which show a very pronounced gradual behavior for both SET and RESET ³⁰⁰.

Furthermore, resistive switching devices with purely electronic switching mechanisms like trapping and de-trapping of electrons at defects states might be promising for analog operation ³¹¹.

6.2. Characterisation techniques

Adnan Mehonic, Wing H Ng, Mark Buckwell, Horatio RJ Cox, Daniel J Mannion, Anthony J Kenyon

6.2.1.Status

Memristive devices pose challenges to the experimentalist both in investigating the physics underpinning device behaviour and in optimising functionality. The wide range of physical phenomena involved in memristance – from metal diffusion in dielectrics to Mott metal-insulator transitions, phase changes and the formation of oxygen vacancy filaments – requires comprehensive physical, electrical and chemical characterisation and even the development of novel analytical techniques ³¹². Here, rather than an exhaustive literature survey we provide examples to illustrate recent progress in characterisation of memristive materials and devices. While we concentrate on oxide-based RRAM materials and devices for reasons of space, most techniques reviewed here are applicable to other memristor types (PCM, MRAM, FeRAM), which require similar characterisation of structural, chemical and electrical changes occurring in in devices as a result of operation. The challenges presented to the experimentalist, particularly when it comes to structural and chemical analysis, are largely similar across all memristive devices.

6.2.2.Challenges and Potential Solutions

Looking first at resistance switching materials and devices, we see that early work from the 1960s on dielectric breakdown suggested the formation of conductive filaments in electrically biased oxide ³¹³ and, while experimental techniques at the time were unable to image them, the authors correctly surmised their existence. The small sizes of these filaments, which can be of the order of a few nanometres in diameter, makes studying their formation and disruption difficult. This is particularly true for oxygen vacancy filaments in oxides; the minimal contrast between the oxide matrix and the oxygen-deficient filament when imaged using electron beam techniques such as TEM-EELS means that there are few direct observations of such filaments. This contrasts with several published TEM studies of metal filaments in oxides, including seminal work by Yang et al  ³¹⁴, which demonstrated field-driven movement of silver ions through SiO₂, Al₂O₃ and amorphous silicon to form dendritic conductive filaments. The large contrast between metal ions and oxide matrix makes it a more tractable problem to image individual conductive filaments. Subsequent work demonstrated different filament growth modes that depend on the relative magnitudes of metal ion mobility and applied field ³¹⁵. Work by Waser et al ³¹⁶ details earlier TEM work on electrical and physical characterisation of resistive switching dielectrics, including observation dating from as early as 1976 of silver dendrites formed in AgS under the application of an external field. Here again, the contrast between the silver filaments and the surrounding matrix provides a significant advantage. It is worth noting that the use of TEM measurements to characterise phase change (PCM) materials and devices is rather easier than in the case of oxide-based RRAM as the contrast between amorphous and crystalline phases of PCM materials is easier to detect, though in-situ and in-operando measurements can pose significant challenges thanks to the fast switching speeds of PCM devices ³¹⁷.

The difficulty of imaging oxygen vacancy filaments using electron beam techniques can be overcome using conductive atomic force microscopy (CAFM) tomography (“scalpel AFM”), a review of which can be found in ³¹⁸. In this technique, sequential CAFM scans of a sample surface imaged using a conductive diamond tip contacting the sample with sufficient force to scrape away the surface provide layer-by-layer conductivity maps that, when stacked, provide three-dimensional images of conductive regions within the oxide. Such studies reveal that electroforming generates large-scale changes beneath device top electrodes, modifying the conductivity of large volumes of material, while one or more highly localised conductive filaments bridge the inter-electrode gap. The technique also reveals details of the internal microstructure of the oxide, showing, for example, the columnar structure of sputter-deposited oxides, as the edges of columns are more conductive than their cores ³¹⁹. It should be noted that this technique only maps conductive regions that are connected to the bottom electrode, so may be better thought of as a measure of connectivity rather than of regions of high conductivity.

The need to apply multiple analysis techniques to probe the movement of oxygen within, and emission of oxygen from, oxides under electrical stress was demonstrated by Mehonic et al ³²⁰. TEM, EELS, CAFM, XPS and mass spectrometry measurements of oxygen emission from samples under electrical stress combine with atomistic modelling to give a fuller picture of the dynamics of oxygen movement and its role in resistance switching. It is clear from such measurements that electrically biasing oxides – particularly those with some structural inhomogeneity – can drive large-scale changes in stoichiometry, reinforcing the CAFM tomography results referred to above. It has been known for some time that this can cause surface distortions and localised bubbling of both electrode and oxide surfaces ^313 321. While work reported in ³²⁰ examines such features using AFM and TEM, the question of how mobile oxygen interacts with electrode materials is partly addressed in recent work by Cox et al ³²². Oxide-based RRAM devices rely on the repeated reduction and oxidation of the switching oxide. To compete with high density Flash, at least 10⁴ cycles are required, and for many applications more than 10⁷ are needed. For such high numbers of cycles, oxygen should not be lost from the switching region around the conductive filament, implying the need for an oxygen reservoir that can both accommodate and release oxygen under appropriate electrical biases. This may be within the oxide, at an oxide/electrode interface, or within one or other electrode. The need to measure oxygen movement is critical. Cox et al ³²² demonstrate that both the electrode material and the microstructure of the oxide layer influence the reversibility of oxygen movement. In the case of electrode metals with high oxygen affinities, oxygen moves easily from the oxide into the electrode, but its movement back again varies significantly between metals. Molybdenum, for example, both accepts and releases oxygen readily when the bias polarity is reversed, while titanium is easily oxidised when positively biased (and when neutral) but does not release oxygen back again when negatively biased. Platinum, having a very low electron affinity, does not accept or release oxygen.

Oxide porosity and sensitivity to moisture is an important factor in resistance switching both by metal diffusion (extrinsic switching) and oxygen vacancy formation (intrinsic switching) ³²³. The presence of moisture in the switching oxide can lead to highly variable resistance switching behaviour. While the origin of such effects remains somewhat unclear, electrical measurements must be interpreted with care in the presence of moisture. The difficulty in measuring hydrogen content in materials and devices reliably and quantitatively is a particular challenge, not only for RRAM, but also other memristive devices. Hydrogen, even in relatively low concentration, can affect the electrical properties of oxides and electrode stacks, as has been recognised for decades in the CMOS community. However, there have been very few studies of the role of hydrogen in memristive devices. At the same time, there is considerable interest in reducing variability and increasing stability in memristors. One cannot help but suppose that detailed studies of hydrogen, and control of its presence, could help these efforts.

Electrical characterisation of memristive devices poses other challenges. A critical issue in RRAM is electroforming conductive filaments. While some materials and devices are forming-free, the majority require an initial conditioning step in which a voltage higher than the normal programming (set or reset) voltages is applied to form a conductive filament that will subsequently be partially oxidised (reset) and reduced (set) by a sequence of voltage sweeps or pulses. Electroforming causes an abrupt drop in oxide resistance, which can span several orders of magnitude, over a timescale of nanoseconds or shorter. Without an appropriate limit on delivered current, this can destroy the device by irreversible oxide breakdown due to Joule heating. It is therefore essential to implement fast current limiting during electroforming. Response times of standard characterisation instruments are generally too long, so current overshoots can over-stress the oxide, leading to breakdown or to conductive filaments too large (hence too strong) to reset. Consequently, full electrical characterisation of devices requires integrated current limiting devices such as transistors or series resistors. Care must be taken to avoid parasitic capacitances or inductances, so the most reliable methods involve on-chip series resistors or transistors in the 1R1R or 1T1R configurations, where the initial R or T labels refer to the integrated current limiter (resistor or transistor), and the second R refers to the resistance switching element.

On the other hand, electrical characterisation of RRAM devices can reveal important clues to the physical mechanisms responsible for resistance switching. Careful analysis of current/voltage curves can indicate a range of electron transport mechanisms, including various forms of tunnelling (direct, trap-assisted, Fowler-Nordheim), thermally assisted transport (Poole-Frenkel) and Ohmic conduction. These, in turn provide evidence for microscopic processes such as charge trapping/detrapping, formation of Schottky barriers or various interface-related electronic states. However, more than one transport mechanism may contribute as, for example, currents may flow in parallel both through a conductive filament and through a highly defective surrounding oxide ³²⁴. In the case of nanoscale filament formation, the thinnest point of the filament can behave as a quantum constriction, allowing only currents that are multiples of the conductance quantum, G₀, to flow. Such effects can be seen easily at room temperature ³²⁴. Reviews of electrical characterisation techniques for resistance switching devices can be found in   ²⁹⁵and ³²⁵.

6.2.3. Concluding Remarks

The inherently interdisciplinary approach that is needed to fully characterise memristive devices poses challenges to the experimentalist. A wide range of techniques are required, which is often beyond the capabilities of a single laboratory, and in some cases these techniques are operating close to their limits. While there have been significant advances in characterisation (CAFM tomography, for example), more work is needed – for instance, to better characterise the role of hydrogen in resistance switching. Where characterisation has been most successful has been when there is close collaboration, not only between experts in different experimental techniques, but also with theorists who provide models to interpret experimental results.

6.3 Comparison between different material systems

Yuchao Yang, Yingming Lu

6.3.1. Status

Varied types of memristors can be realized based on the abundant resistive switching mechanisms in different materials. Each type of memristor has certain characteristics (such as power consumption, switching speed, etc.) that are suited for specific applications, while at the same time, different materials also have their own shortcomings and limitations. A clear understanding of the respective advantages and shortcomings of each material is key to its development and applications.

Among various types of resistive switching materials, transition metal oxides (TMOs) are among the most widely used materials due to their rich resistive switching mechanisms and characteristics. The retention time of TMO-based memristive devices, which indicates how long the resistive state can be maintained after electrical stimulation, is distributed in a wide range from µs to years. According to the retention time, TMO can be generally divided into non-volatile TMO and Mott TMO.

The resistive switching of non-volatile TMO, such as HfO_x and TaO_x, originates from the migration and redox reactions of oxygen ions or vacancies driven by external electric field or thermal effects, which in turn create or destroy conductive filaments between the electrodes. The filaments can exist stably for a long time, and the continuous electrical modulation of geometric characteristics of the filaments such as their lengths or diameters etc., results in multilevel resistive states. Therefore, the non-volatile TMO can be used to imitate the long-term plasticity (LTP) of biological synapses ¹⁷⁰, and can accelerate the computationally intensive matrix-vector multiplication (MVM) in artificial neural networks. Furthermore, due to the mature and CMOS-compatible manufacturing process, a variety of in-memory computing chips based on non-volatile TMO have been demonstrated ⁵⁶. One of the challenges of non-volatile TMO in circuit applications is the existence of the forming process, which requires a high voltage to initialize the TMO layer and subsequently increases the requirements for the voltage and robustness of peripheral circuits. In addition, due to the switching mechanism of non-volatile TMO, the conductance change during the programming process usually shows nonlinear characteristics, along with obvious variations and noises, which increases the difficulty of programming, for example by necessitating closed-loop write-and-verify programming.

There are many other similar TMO systems, such as WO_x and TiO_x. Depending on the robustness of the formed filaments, the retention time can be gradually reduced as a result of filament dissolution, and the conductance of volatile TMO can undergo a continuous decay process. This can effectively map information in time series into high-dimensional vectors, which is widely used in reservoir ¹⁸⁵ for image classification or time series prediction.

Mott TMOs, mainly including VO₂ and NbO₂, show high resistance in body-centred tetragonal (BCT) or monoclinic (M) phases at low temperatures. Once the temperature of Mott TMO exceeds a threshold, the reversible Mott transition occurs and results in structural transition into the rutile (R) phase, accompanied by a significant resistance drop. Based on threshold switching, Mott TMO is widely used in artificial neurons for constructing neuromorphic computing or sensory systems ³²⁶. Meanwhile, the nonlinear transport mechanism of NbO₂ shows high-order complexity during current sweeps, which can effectively realize the rich dynamics of biological neural systems ⁷. However, as the Mott transition is closely related to temperature, the operation of Mott TMO can be affected by ambient temperature. On the other hand, this could serve as the physical foundation for temperature sensing. Furthermore, the metallic domains produced by previous switching events have been found to remain in VO₂ for a long time, which will affect the switching threshold voltage of the devices ³²⁷ subsequently. The transition temperatures of VO₂ and NbO₂ are around 70 and 800 °C, respectively, which are too low and too high from the perspective of circuit applications. A Mott TMO with ideal transition temperature is yet to be developed.

Phase change materials mainly refer to chalcogenide glass materials with reversible phase transition processes. Phase change materials show high resistance in the amorphous state and relatively low resistance in the crystalline state. The phase change memory (PCM) constructed by these materials shows multilevel non-volatile conductance states, and hence PCM can also be exploited to accelerate MVM combined with cross-bar ³²⁸. Traditional phase change materials are the ternary GeSbTe compounds along the pseudo-binary tielines of GeTe–Sb₂Te₃, Ge-Sb₂Te₃ and GeTe-Sb³²⁹. By adjusting the proportions of the three elements in the compound or by incorporating other elements into the compound, these phase change materials can exhibit advantages in various aspects such as on/off ratio, switching speed, and retention. Among these materials, Ge₂Sb₂Te₅, with its outstanding recyclability, is currently widely used, and has been applied in commercial products. In addition to compound-type phase change materials, monatomic Sb exhibits completely different resistances in the crystalline and amorphous states, making it also suitable as a phase change material³³⁰. This type of simplest material can effectively avoid the problem of stoichiometry deviation during the phase change process, which is important for further reducing the size of PCM devices. One of the challenges faced by PCM comes from its conductance drift, caused by the spontaneous structural relaxation in unstable amorphous material, which leads to gradual conductance decay over time and seriously affects computing accuracy and reliability. Therefore, a compensation circuit or strategy for conductance drift is desired. Another challenge originates from the long heating time required for the crystallization of PCM, which not only affects the programming speed ³³¹, but also results in higher energy consumption during programming.

Besides, magnetic materials-based magnetic random access memory (MRAM) is also widely used in neuromorphic computing, which has relatively mature technology, high endurance, etc. The typical structure of MRAM is a magnetic tunnel junction (MTJ), which is composed of two layers (pinned and free) of magnetic material, and an insulator (usually MgO) sandwiched in between.  There are two types of magnetic materials mainly used in MTJ³³². The first one is the multilayers formed by transition metals (e.g. Co, Fe) and noble metals (e.g. Pt, Pd), such as (Co/Pd)_n, (Co/Pt)_n. These materials have advantages in terms of thermal stability and scalability. Another important magnetic material is CoFeB, which shows extremely low programming currents and good matching with the lattice of the MgO barrier layer. By tuning the magnetization orientations of the free magnetic layer to parallel (P) or antiparallel (AP) orientation with the pinned layer, MTJ can exhibit low or high resistance, respectively. Because only the magnetization orientation of the material is changed without large-scale atomic migration or rearrangement, MRAM has low device variation and high reliability. However, the main shortcoming of MRAM is reflected in its relatively low resistance even in the high resistance state ⁷⁶, which will increase the power consumption for MVM computing. Besides, it has a small on/off ratio, which supports only two states, indicating P and AP, and limits its applications in analogue computing.

Similar to MTJ, ferroelectric tunneling junction (FTJ) also changes its resistance by adjusting the polarization orientations of the ferroelectric material sandwiched between two metal electrodes.  The commonly used ferroelectric material systems include perovskite oxides, fluorite ferroelectric materials and wurtzite ferroelectric materials. Perovskite oxides are the most widely used ferroelectric materials with advantages of scalability and switching speed³³³. Among these materials, Pb[Zr_xTi_1-xO₃] (PZT) and Sr₂Bi₂TaO₉ (SBT) have more mature technology and have been applied in commercial memory devices. The fluorite ferroelectric materials, such as HfZrO₂, are also widely utilized and studied due to its CMOS-compatible fabrication process and good scalability³³⁴. The wurtzite materials, such as Al_1-xSc_xN, belong to a new type of ferroelectric material system, characterized by their high on/off ratio, thermal stability, and retention³³⁵. Since the polarization orientations of the ferroelectric layer can be tuned by the electric field with a very low tunneling current, the programming of FTJ consumes low power. Meanwhile, the FTJ shows multilevel conductance ³³⁶ for MVM acceleration in the edge AI platforms. However, ferroelectric materials can have retention degradation caused by intrinsic depolarization fields in the ferroelectric layer, while recent HZO based ferroelectric materials exhibit relatively high coercive voltages and low endurance.

Ion-gated transistor (IGT) is a type of three-terminal device, where the electric field from the gate drives small ions (such as H⁺ and Li⁺) in electrolytes into the device channels to continuously tune conductance of the channel ³³⁷.  Among the reported material systems for IGT, the most chosen material for channel is TMO, which shows the potential for mass production and environmental stability. For the material consideration of electrolyte, the phosphosilicate glass shows clear advantages over previously used Li-ions electrolyte, which is more mature and compatible in the COMS process platform. Owing to the low gate leakage current and separated programming and reading terminals, the IGT shows significantly lower power consumption during programming compared with most two-terminal devices. Furthermore, the weight modulation of IGT can be much more linear, which can greatly reduce the overhead for weight programming compared with devices based on other materials. However, the IGT devices shown to date usually have lower compatibility with standard fabrication processes and have difficulty in the fabrication of large-scale arrays.

6.3.2.Conclusion  

In conclusion, memristors based on resistive switching materials with different mechanisms have demonstrated a variety of encouraging characteristics, such as low power consumption, high scaling potential, fast switching speed, long retention, multilevel conductance, high-order complexity, etc. The MVM engines, artificial neurons, artificial synapses and sensory systems constructed by memristive devicesvi have shown great potential in highly efficient and functional neuromorphic computing. However, depending on the specific resistive switching mechanism, technological maturity and manufacturing cost, there are still technical challenges related to the above material systems. It is important to compensate and correct the adverse effects existing in the application of the materials through the improvement from materials, systems and even algorithms, such as reducing the impact of conductance drift and programming noise on the accuracy of PCM based convolutional neural networks (CNN) by improving the training algorithms ¹⁷⁶. On the other hand, we can also design systems and algorithms to exploit the non-ideal effects in various memristors as resources for improving computational efficiency, such as utilizing the variations and noise in memristor arrays for accelerating the convergence of Hopfield neural networks ²⁰². Once the physical attributes of the memristive devices are properly utilized, they can play important roles in efficient neuromorphic computing.

7. Novel computing concepts

7.1 Embracing variability

Damien Querlioz, Louis Hutin, Elisa Vianello

7.1.1. Status

Emerging nanoelectronic components, such as memristors and spintronic devices, offer exceptional features. However, these devices also exhibit a high degree of variability in their behavior due to their atomic-level features and reliance on sophisticated, sometimes incompletely-understood, physics. This variability has made it necessary to model these devices using statistical tools, effectively treating them as random variables. Interestingly, multiple applications, particularly in machine learning and security, require random variables, which are expensive to generate using traditional CMOS technology. Therefore, exploiting the inherent variability of these nanodevices presents a unique opportunity to develop efficient random number generators and stochastic computing models (see Figure 16).

Figure 16: Illustration of some leading approaches exploiting the variability of nanodevices for computing. Top: measurements of the programming variability of hafnium-oxide filamentary memristors. The statistical rule that this variability naturally implements can be used to perform Markov Chain Monte Carlo training. Bottom: measurement of stochastic magnetic tunnel junctions, naturally implementing p-bits, used, e.g., in Ising machines.  Both devices can also be used for random number genration.

The first idea stems from utilizing the cycle-to-cycle read or programming variability to create random number generators (RNGs) that consume less power than traditional pseudo-random or truly random number generators. By harnessing the inherent variability in these devices, we can develop RNGs that are not only more energy-efficient but also offer improved security and robustness for various applications. Experimental realizations using filamentary memristors ⁹¹, phase change memories  ³³⁸, and spintronic devices ^339,340 have validated this concept. An interesting application is stochastic computing, an alternative approximate low area/low energy computing scheme that has been held back by the lack of compact RNGs: It requires large amounts of RNGs, which in conventional implementations dominate the area of circuits ^341,342.   Stochastic computing can therefore strongly benefit from such stochastic nanodevices ^343,344.

The second idea involves making these RNGs adjustable, i.e., the probability for an output to be one can be controlled by an input signal, effectively creating probabilistic bits or “p-bits” ³⁴⁵. Such structures are analogous to stochastic binary neurons ^346,347, and they have been used in adaptive inference models designed for optimization problems, wherein a set of variables evolves through local and more-or-less random transformations towards configurations that are increasingly probable as they minimize energy ^249,345. A particularly exciting opportunity is their use within Ising machines, which have shown potential for solving highly complex tasks using reduced resources ³⁴⁸. Some of the most promising p-bit implementations use spintronics, as low-energy barrier magnetic tunnel junctions provide a p-bit functionality almost intrinsically  ^249,346,345. Some recent works ^349,350 have used a limited number of stochastic devices as fast high-quality randomness sources for larger FPGA-based circuits, while stressing that the projected benefits of utilizing nanodevices at a larger scale intrinsically harnessing randomness from the thermal bath remain significant and appealing. Memristors with high random telegraph noise ³⁵¹ could also be used in that direction.

The third idea underscores the striking parallels between the behavior of nanodevices and the principles of Markov Chain Monte Carlo (MCMC) algorithms, a class of stochastic optimization techniques. This correspondence is particularly evident in the case of the Metropolis-Hastings MCMC algorithm. In this context, it is required to generate and store multiple random values for a single parameter, a process that can be naturally implemented using the inherent variability of nanodevices. For instance, the programming current can be employed to determine the mean value, while the imperfections of the devices provide the necessary randomness. This family of algorithms has been extensively used for sampling synaptic weight distributions in training Bayesian models, which excel at modeling uncertainty in complex situations. The most important experimental demonstration, using filamentary memristors, is presented in ¹⁹⁹.

While MCMC algorithms serve as a compelling illustration, they are by no means the only example of stochastic computing methods that can benefit from the unique properties of nanodevices. Indeed, a diverse and rapidly evolving field of research is currently exploring the potential of other stochastic computing techniques for a wide range of applications, from neural networks to combinatorial optimization problems. As our understanding of nanodevice behavior continues to deepen, the opportunities to leverage their inherent variability in novel and transformative ways promise to fuel further innovation in stochastic computing and beyond.

7.1.2. Challenges

The primary challenge in exploiting the imperfection of nanodevices lies in the imperfect nature of the imperfections themselves. To harness the inherent variability of these devices for practical applications, it is necessary to achieve a certain level of “controlled” imperfection, which refers to maintaining the desired degree of variability without compromising the reliability and stability of the devices.

The details of memristor  or superparamagnetic tunnel junctions imperfections  are subject to variability. This variability stems from various factors, such as manufacturing variations, environmental conditions, and the complex interplay of atomic-level features and underlying physics. In the case of superparamagnetic tunnel junctions, additional variability in time can occur due to their sensitivity to magnetic field, which is higher than in stable magnetic tunnel junctions. Magnetic shielding can be required to suppress this sensitivity ³⁴⁰. Overall, modeling the imperfections of these nanodevices accurately becomes a challenging task, as capturing these variations in a consistent manner is difficult.

This challenge is further exacerbated by the fact that different types of imperfections have different impacts on the performance and usability of the devices. Therefore, identifying and understanding the specific imperfections that can be harnessed for the development of efficient RNGs and stochastic computing models is of utmost importance.

7.1.3. Potential Solutions

To address the challenges associated with exploiting the imperfections in nanodevices, it is essential to acknowledge that not all imperfections are equally exploitable. For instance, cycle-to-cycle (C2C) variability and device-to-device (D2D) variability differ in terms of their usability and intrinsic nature. C2C variability arises from the inherent variability in the behavior of a single device across different operational cycles, making it more directly applicable and intrinsic to the device. On the other hand, D2D variability occurs due to variations in the performance of different devices, which may be influenced by manufacturing inconsistencies or other external factors.

C2C variability in nanodevices can manifest in two distinct forms: the variability resulting from two consecutive read operations on a device, or the variability arising from two programming operations on the device. Distinguishing between these two effects can be challenging, particularly because both types of variability are highly dependent on the read and programming conditions of the devices. These two forms of C2C variability offer unique advantages from an algorithmic perspective, but they are utilized differently in various applications. For instance, read variability is especially well-suited for RNGs or probabilistic bits ^340,249, while programming variability is specifically tailored for MCMC algorithms ¹⁹⁹.

Even if C2C variability is, in general, more appealing than D2D variability from an algorithmic perspective, D2D variability also has applications: It enables the generation of distinct and irreproducible signatures for each device. Physically unclonable functions (PUFs) are cryptographic primitives that derive their security from the unique and unpredictable physical characteristics of individual devices, and are therefore an excellent example of harnessing D2D variability for practical applications  ³⁵².

Understanding the specific physics underlying different nanodevices can greatly impact their suitability for exploiting imperfections. For example, in hafnium-oxide memristors, only the low-resistance state (LRS) can be readily exploited for our proposed applications, as this state exhibits a much-more controlled degree of variability compared to the high-resistance state (HRS) ¹⁹⁹. Spintronic devices also show promising potential in this regard, as they rely on physical phenomena that inherently exhibit a degree of randomness, which can be understood and modeled ^{340,346,249,353,347}. For example, the switching time of a magnetic tunnel junction is directly connected to the initial angle of the free layers’ magnetization , a random quantity, which can be well-modeled ³⁵⁴, and to some extent controlled. When using low-barrier MTJs as artificial spins in Ising machines, D2D variability may cause an unwanted spread in threshold values or even effective pseudo-temperature across the network. It was shown that these non-ideal and non-uniform activations could be compensated to some degree by relearning the synaptic weights following the initial mapping ³⁵⁵.

Another core idea for dealing with nanodevice imperfections can be to utilize multiple devices instead of relying on a single one. By employing an ensemble of devices, we can achieve better statistical properties, resulting in improved performance and robustness of the RNGs and stochastic computing models. This approach also helps mitigate the impact of individual device variations and reduces the reliance on any single device, thereby enhancing the overall reliability and stability of the systems (examples of this strategy are seen in ^356,357,358).

7.1.4. Concluding remarks

Embracing the inherent variability of emerging nanoelectronic components, such as memristors and spintronic devices, presents a unique opportunity to advance stochastic computing, machine learning, and security applications. By understanding and exploiting the intricate relationship between the variability of these devices and the requirements of various algorithms, we can develop efficient random number generators, p-bits, and Markov Chain Monte Carlo implementations, among other stochastic computing techniques.

However, not every nanodevice imperfection can be exploited. To fully harness the potential of these nanodevices, it is essential to address the challenges associated with their imperfect nature. Achieving a “controlled” level of imperfection, understanding the impact of different types of variability, and leveraging the specific physics underlying each device are crucial steps in this endeavor. By employing multiple devices in an ensemble, we can further enhance the reliability, stability, and performance of the resulting systems. As our understanding of nanodevice behavior and the opportunities to exploit their inherent variability continues to deepen, we can anticipate a surge in innovation in stochastic computing and beyond.

7.2 Spiking-based computing

Sayeed Shafayet Chowdhury, Kaushik Roy

7.2.1. Status

Spiking neural networks (SNNs) are a promising energy efficient alternative to traditional artificial neural networks (ANNs). While ANN based deep learning has achieved tremendous progress in fields such as computer vision and natural language processing, it comes at a cost of huge compute requirements. Spike-based neuromorphic computing ³⁵⁹ provides a potential solution to this issue using brain-inspired event-driven processing. SNNs use  binary spikes for computation contrary to analog values used in ANNs. A schematic of spiking neurons with their temporal dynamics is shown in Fig. 15. The spiking neuron receives spike inputs over time which are accumulated in the membrane potential (V_mem), which upon crossing a threshold (V_th), emits an output spike.

Figure 17: A leaky-integrate-and-fire (LIF) neuron, (a) schematic connection between three pre-neurons to a post-neuron, (b) temporal dynamics of the post-neuron; adapted from ³⁶⁰        

A key characteristic of SNNs is the notion of time.  While conventional feed-forward ANNs are able to map static inputs to outputs with very impressive performance, learning long-term temporal correlations using them is challenging. On the other hand, recurrent neural networks (RNNs) are more suited to process temporal information  efficiently ³⁶⁰; although recent developments such as transformers have shown that ANNs can perform well in temporal processing too ³⁶¹, albeit with a higher training and memory cost. Different variants of RNNs, such as vanilla RNNs ³⁶², long-short term memory networks (LSTMs) ³⁶³, ³⁶⁴ and gated recurrent units (GRUs) ³⁶⁵, ³⁶⁶ have been proposed which differ in their degree of complexity and capability to capture temporal information. However, they all contain explicit feedback connections and memory elements to handle temporal dependencies. On the contrary, SNNs can be regarded as simpler form of RNNs, where the recurrent dynamics of V_mem acts as an internalized memory ³⁶⁷. Interestingly, the leak in SNNs can play the role of a lightweight gating mechanism, thereby temporally filtering out some irrelevant information ³⁶⁸. Additionally, SNNs may lead to lower parameter count and easier training overhead compared to LSTMs ³⁶⁹.

The sequential nature of processing in SNNs leads to unique opportunities in terms of input representation. Traditional feed-forward ANNs such as CNNs and the more recent Vision Transformers ³⁷⁰ process several temporal inputs by merging them into a single large representation. On the other hand, SNNs can process the sequential inputs in a streaming fashion, using the inherent recurrence of neuronal membrane potential. As a result, SNNs are inherently suitable to process temporal event camera ⁴⁰ data. However, for analog inputs, it becomes a key challenge to efficiently encode the data into a spike train. Initial works ³⁷¹, ³⁷² use Poisson rate-coding where the input is compared to a random number at each timestep and a spike is generated if the input is higher than the random number. However, this process suffers from high inference latency. Temporal coding schemes such as phase ³⁷³, burst ³⁷⁴ coding, DCT-encoding ³⁷⁵, time-to-first-spike (TTFS) coding ³⁷⁶, temporal coding ³⁷⁷ etc. attempt to capture the temporal correlation in the data. But their accuracy is often lower than ANNs. More recently, direct encoding approach ³⁷⁸, ³⁷⁹ has become popular where analog values are given directly to the SNN and the first layer of the neural network layer acts as spike generator. Such method has provided impressive performance on complex tasks with very few timesteps. Besides the analog input modalities, DVS cameras (such as DAVIS240  ³⁸⁰) provide discrete spikes directly as inputs which are inherently more amenable to SNNs. Conventional ANN-based approaches tackle the event streams by accumulating them over time and subsequently processing the lumped input altogether ³⁸¹, ³⁸². However, the rich temporal cues present in the event data may not be optimally leveraged in this process. Therefore, recent works propose to process the event streams using SNNs ³⁸³, which leads to a synergistic alliance between the inputs and spike-based processing. We believe this is a promising direction to pursue as it enables harnessing the inherent temporal processing capabilities of SNNs with event data.

7.2.2. Challenges and Potential Solutions

A crucial bottleneck in the advancement of SNNs is the lack of suitable training methods. Sparsity of activations and the discontinuous derivative of spike function leads to training complexity in SNNs. To counter that, initial approaches mostly used ANN-SNN conversion ³⁷¹, ³⁸⁴. Though the conversion method provides high accuracy, it suffers from significant inference latency. Following this, surrogate gradient-based backpropagation (BP) methods have been proposed ³⁸⁵, ³⁷², ³⁸⁶ to train SNNs from scratch. Note, due to the sequential nature of inputs, SNNs are trained with backpropagation-through-time (BPTT), like RNNs. However, simpler neuron models and lower parameter complexity makes the optimization of SNNs simpler compared to RNNs. Although these surrogate gradient-based BPTT methods have advanced the field of SNNs by obtaining high accuracy, the training workloads are still quite intensive in addition to considerable inference latency ( 100 timesteps). To overcome these, the authors in ³⁸⁷ propose to merge the conversion and BP-based training methods (termed as ‘hybrid’ training) where first an ANN is trained to use it as initialization for subsequent surrogate gradient-based BP. More recently, advanced training approaches such as temporal pruning ³⁸⁸, custom regularizers ³⁸⁹ and modified neuron models ³⁹⁰ have been proposed which enable reducing the latency of SNNs to unit timestep. A complementary research direction proposes to utilize equilibrium propagation to train SNNs ³⁹¹, ³⁹². These approaches provide a promising more bio-plausible alternative to backpropagation. However, challenges remain in their large scale implementation.

Parallel to algorithmic developments, advancements in neuromorphic hardware fabrics are equally critical to unearth the true potential of SNNs. Due to the sequential nature of data processing, SNNs present unique challenges on the hardware front as current graphics processing units (GPUs) and tensor processing units (TPUs) are sub-optimal to exploit the high temporal as well as spatial sparsity ³⁹³. Furthermore, information processing using membrane potential over multiple timesteps leads to memory-intensive operations; an overhead that is non-trivial to mitigate using off-the-shelf digital accelerators. Taking such issues into account, several research directions have been pursued in recent years across the stack from devices and circuits to architectures. Event-driven neuromorphic chips such as Neurogrid  ³⁹⁴ and TrueNorth ³⁹⁵ are notable, which are based on mixed signal analog and digital circuits, respectively. Two standout features of these neuromorphic chips are asynchronous address event representation and networks-on-chip (NOCs). Another promising direction is investigating various beyond von Neumann computing models to counter the ‘memory wall bottleneck’. To this end, near-memory and in-memory ³⁹⁵, ³⁹⁶, ³⁹⁷ computing paradigms are being explored to improve throughput and energy efficiency. To realize these emerging computing platforms, exciting progress is being achieved in the device domain utilizing non-volatile technologies ³⁹⁸. Some noteworthy approaches based on memristive technologies include resistive random-access memory (RRAM) ³⁹⁹, phase-change memory (PCM)  ⁴⁰⁰ and spin-transfer torque magnetic random-access memory (STT-MRAM) ⁴⁰¹. RRAMs provide analog programmable resistance but are prone to process and cycle variations and read/write endurance. Devices based on PCM can achieve comparable programming voltages and write speed to RRAMs, however high write-current and resistance drift over time cause issues. On the other hand, compared to RRAMs and PCMs, advantages of spin devices ⁴⁰² are almost unlimited endurance, lower write energy and faster reversal. However, their ON/OFF ratio is much smaller than in PCMs and RRAMs, requiring proper algorithm/hardware co-design ⁴⁰³. Note, each of these technologies has its pros and cons and there is no single winner at the moment. Floating-gate transistors ⁴⁰⁴ are another class of non-volatile devices which are being explored for synaptic storage. While their compatibility with MOS fabrication process is attractive, challenges persist regarding reduced endurance and high programming voltage.

7.2.3. Conclusion

To conclude, SNNs are a promising bio-plausible alternative to conventional deep neural networks. However, it is imperative to understand the ‘why’ and ‘where’ of their proper usage. While SNN algorithms have largely focused on static vision tasks ⁴⁰⁵ till now, their true potential lies in processing sequential information. To that effect, several works are exploring event-based vision for optical flow, depth estimation, egomotion etc. We believe immense opportunities lie ahead in further exploration of SNNs in varied avenues requiring temporal processing such as video processing, reinforcement learning, speech, control etc. In order to achieve that, there is a need of concerted synergistic efforts on algorithms as well as hardware. On the algorithmic aspect, we need to focus on investigating learning approaches that can leverage the unique data representation provided by spiking neurons. Additionally, developing hardware geared towards the SNN specific algorithms is critical for the whole field to move forward. Overall, while domain-specific challenges are prevalent in spike-based computing, we believe the next few years will be exciting as we discover the niche of SNNs, most likely comprising temporal applications with low power requirements.

7.3 Analog computing for linear algebra

Zhong Sun,  Piergiulio Mannocci, Yimao Cai  

7.3.1. Status

Linear algebra problems are being solved in every corner of the information world. Solving these problems by running algorithms in digital computers, however, is generally hard and resource-demanding, featuring a high computational complexity such as O(n³), where n is the number of variables. To overcome the inadequacy of digital computers whose performance is fundamentally limited by the ultimate scaling of Moore’s law and the intrinsic bottleneck of von Neumann architecture, analog computing arises as a promising solution, thanks to its efficient information encoding, massive parallelism, fast response, as well as the emerging resistive memory technology  ⁴⁰⁶. Analog matrix computing (AMC) is conveniently realized with crosspoint resistive memory array, which forms a physical matrix by storing entries as crosspoint device conductances and thus can be used for linear algebra computations. There are several resistive memory device concepts that rely on distinct underlying physics, including two-terminal devices such as resistive random-access memory (RRAM), phase change memory (PCM), magnetoresistive RAM (MRAM), ferroelectric tunnel junction (FTJ) ⁴⁰⁷, and three-terminal devices such as ferroelectric field-effect transistor (FeFET) and electrochemical RAM (ECRAM). They are all simply used as programmable resistive devices to implement AMC ^{40840941041176412}. In this context, one of their differences lies in the conductance range that may limit the capacity of mapping matrix elements, say one bit or multiple bits. It is possible to replace one type of resistive memory device that has been demonstrated for AMC application by another one. For simplicity, we limit our discussion to RRAM that we have used frequently.

The most straightforward AMC implementation is to perform the matrix-vector multiplication (MVM) in one step. By simply applying simultaneously a set of voltages (representing an input vector) to the crosspoint columns, the currents through the crosspoint array are collected at the grounded rows, constituting the output vector which in turn is converted and read out with transimpedance amplifiers (TIAs) (Fig. 1a). By adopting the conductance compensation strategy ⁴¹³, MVM of mixed matrix that contains negative entries can be implemented with the same number of TIAs as in Fig. 1a. MVM is the backbone of many important algorithms, such as neural networks and discrete transformations. Consequently, RRAM-based AMC has widely been considered as an accelerator approach, showing more than two orders of magnitude improvements of throughput and energy efficiency ¹⁷⁹.

In addition to the naive MVM application, more complicated linear algebra computations have been realized through configuring AMC circuits with feedback loops. Figs. 1b-1e show closed-loop AMC circuits for other basic matrix operations, including matrix inversion (INV), generalized inverse, and eigenvector. The INV circuit in Fig. 1b is constructed based on the global feedback connections between crosspoint rows and columns through operational amplifiers (OPAs). It solves a system of linear equations when an input current vector is provided, with the output voltages of OPAs representing the solution ⁴¹⁴. INV is exactly the inverse problem of MVM, both circuits utilize the same electronic components, whereas the different connection topologies define the opposite functions. The INV concept can be generalized to non-square matrices, by configuring AMC circuits with two sets of crosspoint arrays and amplifiers. Depending on the matrix shape (tall or broad), the generalized left (Fig. 1c) and right (Fig 1d) inverse circuits have been designed, both based on the same feedback loops while showing differences in terms of input terminals and matrix storage (transpose or not) ¹⁹². INV and generalized inverse find applications in many scenarios, such as machine learning, wireless communications, and scientific computing ^415416417. Unlike MVM, INV suffers from the condition number issue, where the input error would be amplified for an ill-conditioned matrix (with a large condition number), resulting in a low computing precision. Consequently, it is appropriate to use the INV circuit as an analog preconditioner, and the performance of scientific computing may be improved by more than three orders of magnitude ⁴¹⁵. The eigenvector circuit in Fig. 1e uses global feedback as well, but it is a fully self-sustained system with no external inputs, while working by positive feedback mechanism ⁴¹⁴. It finds applications in typical scenarios including quantum simulations, PageRank for Google search or recommender systems.

Figure 18: AMC circuits for (a) matrix-vector multiplication, (b) matrix inversion, (c) generalized left inverse, (d) generalized right inverse, and (e) eigenvector computations.  

Recently, more AMC circuits have been developed for solving more complicated matrix problems. Fig. 2a shows a design for matrix-matrix-vector multiplication (MMVM), by mapping two matrices (or two copies of one matrix) in a RRAM array, and assisted by the use of conductance compensation ⁴¹⁸. By connecting this MMVM circuit with other analog components, and particularly a nonlinear function module based on operational amplifier, to form a feedback loop, the resulting circuit solves the sparse approximation problem in Fig. 2b in one step without discrete iterations. Notably, the nonlinear function may also be implemented by a volatile resistive switching device, thus substantially improving the compactness of the AMC circuit ⁴¹⁹. It has been used for compressed sensing recovery pf sparse signal, natural and medical images, representing a highly promising solution for the back-end processor to deliver real-time processing capability in the microsecond regime.

Figure 19: AMC circuits for (a) matrix-matrix-vector multiplication, and (b) solving sparse approximation problems, e.g., compressed sensing recovery.

7.3.2. Challenges

Thanks to the manufacturability and compatibility of RRAM device in modern CMOS technology process, large-scale RRAM arrays have been fabricated for AMC implementations. Particularly, RRAM macros including peripheral circuitries are usually designed to deliver MVM accelerations. By contrast, closed-loop AMC has been limited to small-scale concept demonstration, e.g., for 3×3 matrices. The reason behind the developmental stagnation might be ascribed to the unconventional analog circuitry. Different from MVM that consists of only local feedback, closed-loop AMC circuits contain complicated, hard-wired global feedback connections across the entire RRAM array. Since all elements are involved to provide a collective circuit response, the operation may become sensitive, risking the damage of RRAM devices under excessive electric stimulus. In addition, for large-scale circuits, the non-ideal factors of devices and circuits will jointly lead to an exaggerated deviation from the correct result. Also, the time response of the circuit might be influenced by non-idealities such as parasitic resistances and capacitances, which may even cause an instability issue.

Despite the lack of large-scale demonstration at this moment, it is quite promising to build closed-loop AMC circuits based on the relatively mature RRAM technology in the near term, given that lots of theoretical and simulation works have intensively examined the potential issues. In the long term, the following aspects shall be addressed to support the development of AMC circuits for linear algebra:

(1) Device/array level: At present, the largest available array size is 512×512 ¹⁷⁴ . In practice, only a fraction of the array is turned on for computation, due to the current/power overload and the accuracy limitation. We believe such an array size is already sufficient to well support the advantages of AMC over other paradigms. Otherwise, we should put the stress on the analog conductance tunability of RRAM devices, which are expected to show as many distinguishable conductance levels as possible. In turn, fast and accurate programming of the RRAM conductance is essential to maximizing the AMC efficiency. The linearity and symmetry of device conductance update has been continuously emphasized for the online training of neural network weighs ⁴²⁰, such a characteristic should also be favored for real-time update of matrix elements in general AMC applications. On the other hand, there should be a tradeoff consideration on the RRAM conductance range, which is associated with the power consumption, alleviation or exacerbation of the impacts of resistive and capacitive parasitics. The device variations should matter most to affect the computing precision. In particular, for solving inverse problems with the closed-loop circuits, the matrix structure related to the condition number should also be a deterministic factor.

(2) Circuit/architecture level: The scalability of AMC circuits comes at the price of the reduced precision, where the accumulated error may eventually decline the nominal result. As a result, there is a tradeoff between the desired computing accuracy and the possible array size. For solving extremely large-scale problems, especially for the inverse matrix problems, it is imperative to have algorithmic solutions to recover the result correctly. Therefore, clever design of efficient algorithms for matrix processing would be very helpful. Particularly, as RRAM-based AMC uses two physical attributes, namely conductance and voltage, the cascading of an algorithm may require the transition between the two attributes, which will inevitably make the operation complicated. To this end, algorithms featuring as less such transitions as possible are precious. In addition, since AMC circuits are all based on the core RRAM array, it is valuable to have a reconfigurable architecture for performing different matrix operations. As amplifiers play another critical role in AMC, it is highly beneficial to design efficient amplifiers that are well suited to the circuits.

(3) Software & application: During the AMC operation, multiple rows/columns of the array are simultaneously activated to carry out computation, indicating a primary requirement of such instructions in an AMC program. Additionally, instructions for circuit reconfigurations should be included to enable a general AMC architecture. Therefore, the efforts on compiling and programming for AMC will be of significant importance  ⁴²¹ . To support the AMC concept, finding a killer application that well matches the advantages of AMC will be most convincing. The application of MVM to neural network acceleration has attracted enormous attention in past years, with a few successful silicon demonstrations towards real-world applications. For closed-loop AMC circuits, we believe non-linear matrix problems and related applications would be most promising, since they are more tolerant to errors as in the neural network.

 7.3.3. Potential Solutions

RRAM-based AMC is extremely fast in that the computation is basically a parallel reading process. To prevent the memory writing process being a bottleneck, parallel writing schemes have been developed. For instance, the weight matrix in the array can be updated by using the outer product operations, sometimes assisted by gradient decomposition methods ⁴²². However, these methods usually overlooked the memory nature of RRAM, that is, the analog information needs to be reliably read out. To this end, a verification circuit should be included to confine the conductance distributions. In particular, fast writing with less/none iterations will be very beneficial to saving the latency. One example is the closed-loop write scheme that uses current feedback to control the resistive switching process ⁴²³. Because of the underlying ionic migration mechanism of RRAM devices, the endurance capability has constantly been a critical issue for both memory and AMC applications ⁶². It is unlikely that this issue can be overcome solely by test method innovations, new physical mechanisms may be needed to solve it from the source. Additionally, investigations on device materials and structures to fundamentally optimize analog RRAM performance and to empower the array extension are always highly desired.

The scale-accuracy tradeoff of RRAM-based AMC should be elaborated to balance the efficiency and reliability. Efficient algorithms for matrix tiling and result recovery are vital for solving large-scale problems. Recently, a scalable AMC method, termed BlockAMC, has been proposed for solving large-scale linear systems. It partitions a large original matrix into smaller ones on different memory arrays, and performs MVM and INV operations with the block matrices to recover the original solution ⁴²⁴. To reduce the impact of non-idealities, AMC systems with both algorithmic and architectural innovations should be designed. For MVM, it is convenient to implement the bit slicing method to extend the computing precision, by using only low-precision memory devices ⁴²⁵. However, as the INV circuits contain global feedback loops, it is difficult to apply this method to improve the precision of INV operations. Eventually, an analog-digital hybrid system may be required to deliver this capability. With the RRAM array as the core, the peripheral circuits and the connections can be reconfigured to perform different AMC operations. In this regard, the basic models shall include RRAM array, amplifiers, and reconfigurable routing, etc. Different circuit configurations are actually hardware-embedded instructions of matrix operations.

It is of great practical convenience to have a set of fixed AMC primitives that can be adopted to realize general matrix computations, together with vector processing, parameter scaling, and variable attribute conversion. There is a trend that takes MVM and INV as two primitives to enable a reconfigurable, general-purpose AMC system, which may be used for linear regression, generalized regression and eigen-decomposition ^426427428. Given the intrinsic noises in analog computing, we believe non-linear closed-loop AMC that are inherent in error tolerance are more promising towards real-world applications. Typical examples include solving some optimization problems, such as sparse coding, compressed sensing recovery, as well as linear/quadratic programming. These problems appear in many common scenarios such as wireless channel estimation, magnetic resonance imaging, and signal processing  ⁴²⁹ . Also, front-end integration with sensors for signal processing would be encouraging, which helps save data conversions and thus reduces the accumulation of noise effects  ⁴³⁰ . These problems typically favor relatively small RRAM arrays, thus alleviating the rigorous requirements on the device/array performance. Eventually, as the RRAM technology matures, the application to large-scale and high-precision problems will be advanced, with the help of innovative algorithms and architectures.

 7.3.4. Conclusion

In the modern era, due to the strong demand for linear algebra acceleration and the rapid development of emerging resistive memory devices/architecture, analog computing has gained a renewed interest across academia and the industry. Various AMC circuits have been successfully demonstrated for fast solutions of matrix problems that constitute the basic operations for linear algebra computations. However, most of the AMC concepts still remain in the laboratory prototype stage, calling for a roadmap covering different aspects to guide system integration, optimization, and application. Next steps towards effective AMC shall include developing (1) reliable analog conductance programming methods, (2) architecture and algorithm designs for large-scale problems, (3) circuit designs for non-ideality mitigation, (4) reconfigurable systems, (5) more AMC circuits, e.g., for non-linear matrix operations, and (6) application to near-term and long-term typical scenarios.

7.4 Analog Content Addressable Memories (CAMs) for in-memory computing

Giacomo Pedretti                           

7.4.1. Status

Content addressable memories (CAMs) are a class of memory structure which given an input query returns its stored location, or address. A wildcard, ‘X’, can be added allowing for ‘fuzzy’ searches, resulting in a ternary CAM (TCAM)  ⁴³¹.

Figure 20: Illustration of a TCAM (a) and an analog CAM (b). (c) Circuit schematic of a 6T2M analog CAM, able of returning a match if the input on the DL is within a function of the memristor condutance (inset). (d) Example of decision tree and (e) its mapping to an in-memory computing circuit for single/few cycle inference.

Figure 20(a) shows a TCAM schematic, an input query is applied along the columns, or search lines (SLs), and outputs are returned along the rows, or match lines (MLs). A given row is matched (returning a ‘1’) if each value $q_{j\ }$of $q$ is equal to the corresponding key-value $K_{ij}$ stored in the row $i$(column $j$), which corresponds to performing the operation

$$ML_{i}\ =\ \prod_{j}\overline{\left(q_{j}\oplus K_{ij\ }\right)}.$$

CMOS/SRAM-based TCAMs are ubiquitous in networking, but they are usually bulky and power-hungry. Memristor-based TCAMs have been shown to outperform the CMOS circuits but given the relatively low speed and inefficient writing of memristors, they are currently being proposed for in-memory computing applications.  In this framework, recently we proposed an analog CAM  ⁴³² which stores ranges and returns a match on the MLs if the analog input is within the stored ranges or

$$ML_{i}=\prod_{j}\left(T_{l,ij}\leq q_{j\ }<T_{h,lj\ }\right)$$

where $T_{l}$ and $T_{h}$ are the stored upper and lower bounds thresholds, respectively.  Figure 20(c) shows a conceptual representation of an analog CAM and  Figure 20(c) the schematic of a 6-transistors-2-memristors (6T2M) analog CAM. Two 1T1R voltage dividers compare the analog input on their gate with the stored value in the memristor conductance, and their drain node controls either a pull-down transistor (left side) or a series of an inverter and a pull down transistor (right side) to discharge the pre-charged ML if the input is lower than the lower bound or higher than the upper bound, respectively. For example, if the input voltage is high enough to turn on the input transistor resulting in a voltage on the left-most 1T1R drain lower than the pulldown threshold, the latter won’t be activated. Similarly, if  the same input voltage is low enough such that the input voltage of the inverter is above its threshold of the right-most 1T1R, its pulldown transistor would not be activated. This results in a match, given that the ML would not discharge.

Other implementations of analog CAMs have been proposed based on compact ferroelectric structures, although given the absence of the inverter, a proper conditioning of the input should be performed ⁴³³.

While the in-memory computing community in the last decade  have been focused on the acceleration of deep neural networks (DNN) (as presented in paragraph 4.1 SRAM and 5. Emerging Technologies (computing approaches)), tree based machine learning (ML) still outperforms DNN when processing tabular data due mainly to the presence of missing and categorical features ⁴³⁴.   Figure 20(d)  shows a schematic representation of a decision tree (DT), which is essentially a collection of conditional branches (nodes) in a tree structure. Given the irregular structure, ensembles of DTs are not well suited for being accelerated in CPU and GPU, due to thread synchronization issues, load imbalance and uncoalesced memory acces  ⁴³⁵.  Recently, we showed that DTs ensembles can be mapped to analog CAM, providing a one-cycle and conversion-less inference, given that the output is already digital and ready to be post processed  ⁴³⁶.  Figure 20(e) shows the DT of Figure 20(d) mapped into an analog CAM array, where each root-to-leaf path is encoded in a row. The digital MLs output are connected to a conventional RAM storing the leaves values, or prediction. All branches are executed in parallel, providing a size-independent inference latency.

7.4.2. Challenges

Most of the challenges for building reliable analog CAM are shared with crossbar arrays of memristors, although some circuit and device requirements are different. Similarly, a wide range of programmable levels in the memristor conductance is desirable, but different from crossbar array slicing techniques for increasing precision ⁴³⁷ can’t be performed. Range-based analog CAMs have a unary encoding thus connecting multiple adjacent analog CAMs on a given row doubles the number of levels as opposed to bit-slicing in crossbar arrays which doubles the number of bits, leading to an exponential overhead. Linear approaches for improved analog CAM precision have been recently shown ⁴³⁸, although with significant circuit complexity overhead. Moreover, given the voltage divider operation performed directly on the memristor, its conductance needs to be stable at higher read voltages, e.g. 1 V, to be able to control efficiently the pulldown transistor.

While a small 2x2 array was realized in the 180 nm technology node and large array simulations at the 16 nm technology node were performed ⁴³² experiments on large arrays, i.e. 128x16, have yet to be realized. There are several concerns to be solved at the array level, for example, the impact of memristor variation and noise on the search accuracy and the impact of parasitics capacitance and resistances on a reliable operation. Moreover, due to the non-linear input/output relationship, even defining the appropriate patterns for the program is challenging and ad hoc circuit-aware programming routines should be performed.

Finally, a full and general architecture design has yet to be performed, to feed and pull data with enough bandwidth to take advantage of the single-cycle memory lookup operation, eventually including multiple independent cores and performing operations on a large amount of data in parallel. The architecture should be programmable, not limited to tree-based ML inference,  allowing multiple tasks such as associative searches ⁴³³, and resilient to errors due to memristor variations for example including error correction codes routines.

7.4.3. Potential Solutions

First, large arrays of CMOS-integrated analog CAM with memristor devices should be designed and fabricated. Techniques for increasing the memristor conductance stability at relatively high reading voltage have been presented and should be used to develop analog CAM-specific memristor stack. For example, a larger oxidation layer or/and a controlled deposition to limit the oxygen defect in it can be used to increase the set voltage of memristor devices, which could result in better stability.

Analog CAM circuits to support a higher number of bits can be designed by studying the appropriate logic functionalities in the case of a separate comparison of the least significant and most significant bits, for example by mapping different logic functions between adjacent cells ⁴³⁹. In principle, it is possible to implement any kind of logic operation between adjacent CAM cells by opportunely connecting the pulldown transistors. Custom program and verify algorithms, efficiently including process variation and noise are being developed ⁴⁴⁰ , in order to design appropriate target patterns to program in the memristor conductance while performing non-linear operations.

Finally, a similar architecture to the crossbar array accelerators with a custom instruction set architecture (ISA) compiled from popular tools such as sk-learn, can be implemented ⁴⁴¹. The analog CAM operation should be abstracted enough in order to efficeintly map different workloads in a hierarchical way to the programmable accelerator, with multiple cores handling a different part of the problems, and a global network on chip accumulating the results to finalize the computation. In order to make each core operation reliable, an error detection scheme already designed for TCAMs can be adapted to the new analog CAM to efficiently re-program a given device once the state has been drifted out of the desired one. ⁴⁴²

7.4.4. Conclusion  

While a lot of attention has been spent in the last years by academia and industry research for developing in memory computing structures based on linear operators, such as crosspoint arrays, recently a renewed interest in memristor-based CAMs has arisen. Different circuit for memristive TCAMs and analog CAM have been proposed, targeting multiple applications, but a circuit-level integration of CMOS circuitry with a BEOL-integrated memristor device along with a system-level analysis of performance at scale has yet to be shown. While engineering such required milestones, researchers should also focus on exploring new applications exploiting open-sourced circuit models, given the novelty of the idea it is in fact likely that we are just scratching the surface of the potentiality of such computing primitives given the inherited non-linear operation performed with unprecedented speed and energy efficiency.

7.5 Optimization solvers 

John Paul Strachan, Dmitri Strukov

7.5.1. Status

One of the best examples of a high-risk-high-reward area for novel computing is in the area of optimization solvers.  Here, the ambition is to offer some form of speed-up or reduced resource requirements (memory, energy, etc.) in solving computationally expensive combinatorial optimization problems.  In such problems, the task is to minimize a given cost (or energy) function by choosing the best configuration within a large dimensional space.  Well-known examples include the Traveling Salesmen problem (finding a minimal route uniquely visiting various cities), graph coloring (coloring nodes such that connected nodes have different colors, while using the fewest colors ), or training the weights in an artificial neural network (ANN).  While some problem classes and instances can be solved approximately and quickly using greedy algorithms, many remain intractable and no known algorithm exists that can solve them all with polynomial resources and in polynomial time.  Even a modest speed-up could offer immediate benefits in practical applications across planning, wireless communications, bioinformatics, routing, finance, and many more.

The above challenge has inspired many research communities to develop both new algorithms and new physical hardware approaches, often in conjunction.  A prominent example of this was the recognition that many optimization problems can be mapped to problems in physics⁴⁴³, such as finding the ground state of an Ising system of coupled up-or-down spins. This has driven interest in so-called Ising solvers, which can have many physical realizations including optical parametric oscillators, coupled electrical oscillators, magnetic tunnel junctions, analog memristors, and others  ³⁴⁸.  From the early neural-inspired communities, John Hopfield’s nonlinear networks of analog neurons turned out to also solve planted optimization problems⁴⁴⁴. Quantum or quantum-inspired approaches have also been pursued, leveraging the quantum adiabatic theorem, where a ground state (optimal) solution is found by evolving slowly from an initial, simple Hamiltonian toward a final Hamiltonian that encodes the desired problem⁴⁴⁵.   Dynamical solvers are also being pursued⁴⁴⁶⁴⁴⁷, where the large configuration space is explored rapidly via coupled nonlinear dynamical equations, with optimal configurations acting as attractor states.

It is important to realize that the variety of physics and quantum-inspired approaches discussed above offer primarily an algorithmic paradigm for solving optimization problems.  The underlying materials, physics, and computational elements remain quite flexible.  For example, traditional CPUs, GPUs, and FPGAs can be used to simulate nonlinear dynamical equations.  Or electronic CMOS-based ring oscillators can be built and resistively or capacitively coupled in order to simulate the magnetically interacting spins in an Ising model.  Thus, there are many possible combinations of algorithmic or computational models and the physical substrate where computations are performed. It is easy to mistake one for the other.  Yet, despite such variety in possible physical realizations, some common underlying challenges emerge that must be handled in order to successfully engineer an efficient and broadly useful optimization solver.  The remaining portion of this document lists these challenges and opportunities.

7.5.2. Challenges

Challenges of mapping to hardware 

Mapping optimization problems to physical systems, such as a hardware Ising model (illustrated in Figure 21), is an attractive approach, but also highlights many key challenges.  An Ising model is based on pair-wise couplings of binary spins, and therefore, up to quadratic terms appear in an energy function with binary variables, and there is no explicit mechanism to enforce additional constraints (such as summing to certain integer values). This is also known as a Quadratic Unconstrained Binary Optimization (QUBO) type, having an energy function in the case of an Ising model,

$$E_{\textrm{QUBO}}=\frac{1}{2}\sum_{i,j\neq i}W_{ij}s_{i}s_{j}+\sum_{i}h_{i}s_{i}\ .$$

Yet many optimization problems involve higher than quadratic interactions.  For example, a k-SAT Boolean satisfiability problem involves terms of order k. Such higher-order interactions must be mapped down to quadratic terms, through the introduction of additional auxiliary variables.  The result is a new total number of variables that is polynomially larger than the original number of variables, leading to, in the worst case, an exponential penalty in terms of the configuration space that needs to be searched. This has motivated explorations of algorithms with higher-order interactions⁴⁴⁸ and their physical realizations.

Figure 21: Typical steps for solving combinatorial optimization problems with Ising machine. The original problem is formulated as quadratic unconstrained optimization (QUBO) problem. The QUBO problem may need to be modified, e.g., sparsified or decomposed to smaller subproblems, before it can be mapped on the targeted Ising machine. Note that the figure shows fully connected Ising machine and simplest baseline QUBO formulation. Efficient implementation of different annealing techniques approaches and solving constrained optimization problems relies on the ability to adjust coupling weights during runtime.

Additionally, QUBO formulations are sometimes inefficient, even for natively quadratic problems. For example, the typical QUBO approach for a K-city traveling salesman problem is to encode each route with K one-hot-encoded K-bit vectors representing the visitation order of each city ⁴⁴³.  The result is a quadratic scaling for QUBO variables, i.e., N = K², with the number of cities. This can lead to a worst-case exponential penalty in the configuration space to search as a function of the number of variables.

Scaling challenges

Efficient mapping (“embedding”) of an optimization problem to the underlying Ising hardware may require further modification of the QUBO formulation.  Indeed, the naive implementation of N-variable QUBO problems requires N² coupling weights to allow programmable coupling with unique weights between any pair of neurons. N can be more than a million for many practical sparsely-coupled optimization problems (such as already mentioned TSP and SAT problems), which is clearly unacceptable for most Ising model implementations. To address this challenge, a general approach is to decompose the original QUBO problem into smaller subproblems that can be implemented in the hardware. For example, one can partition a large neuron connectivity graph into smaller subgraphs, with minimized number of edges between subgraphs. The corresponding subgraph subproblems are then solved independently, e.g., by fixing values of variables not participating in the currently solved subgraph. The downside of such an approach can be lower solution quality or longer time, even if all subproblems are individually solved optimally. The scaling problem is further exacerbated for hardware approaches with limited coupling capabilities.  For example, if only limited neighbor connectivity is possible – such as in the so-called Chimera graph topology implemented in the DWave interconnected superconducting bits  – then the embedding comes with even higher overhead. In this case, the original QUBO problem is first sparsified to meet the connectivity limitations by adding redundant variables, and then partitioned into subproblems. The result can be an exponential slowdown⁴⁴⁹ in solution convergence.

Precision challenges

Many emerging device technologies do not allow for realizing accurate weight couplings and performing precise computation of node updates. On the other hand, the weight precision should be sufficient to encode the weight dynamic range. More generally, the non-idealities in the physical couplings need to be low enough not to distort the global energy minima. Detailed studies of non-idealities in, for example, Ising machine operations are so far very limited; however, an initial insight is provided by similar studies for training artificial neural networks, a type of optimization task. Training to the highest accuracy requires at least 4-bit weight and dot-product computation precision for many deep learning models⁴⁵⁰. The precision requirements are higher for very compact models such as MobileNet and lower for larger, redundant models such as VGG. This is likely inversely correlated with the number of global minima and their basin volumes. Therefore, higher precision requirements are expected for solving harder combinatorial optimization problems and/or when better solution quality is sought.

7.5.3. Potential Solutions

Advances are needed across algorithmic, device, and architectural levels.  On the algorithmic front, it was already stressed above that improved mappings are needed to ensure efficient conversion from the original optimization problem to QUBO, as well as efficient embedding.

On the device front, we need to identify the best device and material implementations to overcome challenges in connectivity and interactions. Are there any material and device concepts that would allow implementation of the higher order in the hardware? The scaling issues highlight the importance of efficiently implementing coupling weights. In this respect, emerging analog memory devices and in-memory computing with dot-product circuits are especially attractive. Optical computing approaches are attractive because of the high fan-in interconnect that can be attained. Quantum computing is attractive because of quantum annealing, which offers a tunneling mechanism through energy barriers.  Many tricks on the device and circuit levels can be borrowed from the work on neuromorphic inference to improve tolerance to non-idealities and increase precision.

On the architectural side, we need more flexible designs to handle broadly varying levels of difficulty and size. Many problems are locally dense but globally sparse. For example, hard k-SAT problems are sparse, with sparseness increasing with model size, and the architecture should take advantage of such sparseness by mapping highly interconnected neuron subgraphs onto fully connected smaller size Ising machines and interconnecting Ising machines with routing networks.

A key feature for further advances on every front will be proper benchmarking. The design space is very broad and is further complicated by the wide range of applications with different characteristics (hardness, coupling density, problem size). Figure 22 shows the key metrics – solution quality found as a function of hardware time to solution at a fixed power consumption and hardware resource budgets, highlighting the important tradeoffs to identify promising solutions and guide further design. Such a figure can be drawn for a collection of benchmarks or an individual benchmark and, in the general case, would change with the scale of the problems.  Note that the fixed resource budget in Figure 22 is essential for a fair comparison of different designs because time to solution can be improved with parallelism, i.e., spending more energy and/or relying on higher complexity hardware. (Alternatively, Figure 22 can be extended to show time, energy, and hardware complexity to the solution quality. Such characterization would be more complete, though less insightful when sparse data are available.)

Figure 22: Benchmarking and design space for optimization hardware (shown schematically). We expect solver implementations of type (A) that are suitable for quickly finding lower-quality solutions but incapable of reaching higher solution qualities, such as, for example, fast but approximate hardware heavily relying on parallel decomposition algorithms and/or imprecise circuits. At the other end of the spectrum are brute-force (complete) approaches (C), e.g., running an algorithm of checking all possible solutions on the conventional high-precision computer, that are slow but guaranteed to find the global optimum. Other approaches (B) may perform better for medium solution quality and time cost. We expect a Pareto front formed by different hardware approaches highlighting that linear improvement in solution comes at exponentially longer hardware times. The figure is motivated by similar dependencies at the algorithmic level (so-called run-length distribution figures) and similar hardware tradeoffs in neuromorphic inference and training ⁴⁵¹.  

For example, we anticipate that low precision and/or restricted connectivity systems will be much faster and energy-efficient when worse precision is required. Such systems may never reach the highest quality solutions, e.g., because inherent error modifies the energy landscape or the subspace. On the other hand, a brute force algorithm, e.g., performed on a conventional computer, would guarantee finding optimal solutions. The first emerging optimization solvers (such as Ising machines) can occupy intermediate locations in these trade-off curves, tuned to the desired priority metrics.

7.5.4. Conclusions

There is a great potential to harness physics- and brain-inspired approaches to improve today’s computing systems for solving optimization problems.  Yet, we see challenges that must be overcome at many materials, circuits, system, and algorithmic levels in order to realize this potential.  A variety of approaches are under exploration that use optical, quantum, magnetic, or electronic components, sometimes in combinations.  But the core requirements and issues are similar across all of them, as outlined here.  In the end, it will be critical to engage other, non-hardware communities to become enthusiastic users and to help tool development for these emerging hardware systems. It will only be through continuous feedback between the users, developers, material scientists, and engineers that a reliable and flexible optimization solver can be developed for large-scale practical problems.

8. Technological Maturity

8.1. Current status & next steps

Manuel Le Gallo, Stefano Ambrogio

8.1.1. Status

The various neuromorphic computing technologies and concepts covered in this roadmap have shown promising results through various prototype chip demonstrations realized by academic and industrial institutions. SRAM-based in-memory computing accelerator chips could demonstrate energy efficiencies >100 TOPS/W for 4-bit matrix-vector multiplications and provide support for all the additional CNN and LSTM inference operations with on-chip digital ALUs ⁴⁵². The recent analog accelerator based on nor-Flash memory from Mythic, supporting 80M on-chip analog weights, demonstrated 3.3 TOPS/W system-level energy efficiency for 8-bit calculation precision and could run a pose-detection application while consuming only 3.73W ⁴⁵³. While several alternative technologies have been recently explored, such as ferroelectric and magnetic devices, 2D transistors and memtransistors (see Section 3.2 Different neuromorphic technologies and state of the art), the investigations are still at the level of single device or small array, preventing a proper computation performance exploration. For this reason, in non-CMOS based implementations, only the more mature PCM and RRAM emerging technologies have been integrated into multi-core in-memory computing chips and could demonstrate various neural network inference tasks, albeit not fully end-to-end. Near software-equivalent accuracies and energy efficiencies of 10 TOPS/W or higher for matrix-vector multiplications have been reported ^{454,174,56,455}. Less mature technologies such as ferroelectric and magnetic memories have been successfully integrated into small arrays with on-chip data converters ⁷⁶.  Memories based on spintronic, 2D and atomistic materials as well as photonic processors based on resistive memories have been mainly investigated at the individual device level or integrated into small arrays without peripheral circuits performing data conversions ^269,282.

Besides chips that aim at accelerating matrix-vector multiplications for deep neural network inference tasks, platforms that can execute more novel neuromorphic computing concepts via in-memory computing have been demonstrated as well. A 64k-cell PCM chip from IBM with in-situ learning capability using STPD and leaky integrate-and-fire neurons performing a simple associative learning task was demonstrated ⁴⁵⁶. Another 1.4M-cell PCM chip implementing a restricted Boltzmann machine with STDP learning rule could demonstrate low-power on-chip training and inference on the MNIST dataset ¹⁷¹. Spiking implementations are also used to efficiently implement ResNET networks on CIFAR-10 ⁴⁵⁷. Although those platforms are small prototypes and do not support end-to-end deployment of a variety of models, the key computational blocks involved in the execution of the algorithms have been successfully integrated on-chip together with the in-memory computing crossbar arrays. Nonetheless, such prototype demonstrations are still far behind the maturity of digital CMOS computing platforms, which support full deployment of a wide variety of models, often with end-to-end software stack. CMOS neuromorphic computing chips such as IBM TrueNorth ⁵⁰ and Intel Loihi ⁵¹ have been made available as research platforms to implement inference and training of spiking neural networks with some software support being provided. Moreover, several digital application-specific deep learning accelerators with end-to-end software stack, such as Google Tensor Processing Unit, Amazon Inferentia, Facebook’s M.2 accelerator, and IBM Artificial Intelligence Unit have reached a mature state in terms of production and system-level integration ⁴⁵⁸.

8.1.2.Challenges

Several challenges at the device- and system-level could hinder the development of fully end-to-end in-memory computing accelerators. The growing interest in inference tasks, where the neural network is firstly trained in software and then deployed on-chip to get high throughput (number of inferences per second) and low latency (time to process one input) requires chips with multiple crossbar arrays, to account for multiple layers in neural networks, and reasonably large size, to map extended network layers (e.g., first layers in convolutional networks) ^454,459. This leads to several device-level requirements, such as high HRS/LRS resistance ratio, moderate endurance, high retention, and low intrinsic variability. It is also beneficial to have a LRS resistance high enough to limit the impact of the voltage drop in the lines of the crossbar array during writing and readout. Although a single device can be designed to easily meet one of those requirements, the challenge is to build multiple chips each containing multiple arrays of devices that should meet all of them. Because it is rather challenging to simultaneously achieve such specifications with emerging technologies, often tradeoffs have to be made depending on the envisaged application. As an example, a lower endurance can be acceptable to tradeoff for a higher HRS/LRS ratio for inference purposes. In addition, several aspects regarding the operating voltages and currents of these devices need to be considered when integrating them into crossbar arrays. Programming voltages in excess of 2V and large programming currents (>100µA) necessitate the use of large access transistors that can block this voltage and drive this current. Therefore, minimizing the programming voltages and currents is critical to achieve high device density in the crossbar.

On training chips, where the weights of the neural network are actively changed on-chip to get high-accuracy processing on several tasks (classification, language processing, …) specifications are higher. In addition to the inference requirements, data needs to back-propagate through the network. Such behavior is critical in analog cores, requiring fully symmetric peripheral circuitry during forward and back propagation, absence of any polarity dependence of the resistive devices (since each memory element is generally biased in a different way between forward and back propagation). Clearly, device challenges are different: while inference requires high retention, low drift and low temperature dependence, training requires high endurance and symmetric conductance update behavior, leading to different material choices.

Moreover, to achieve highly competitive system-level performance against existing CMOS-based accelerators, further improvements in device-level precision and compute density are required ⁴⁶⁰. While this poses material challenges on resistive memories, it also influences the peripheral circuitry by limiting its available area while ensuring that it does not restrain precision. In addition, real workloads involve a variety of different operations other than matrix-vector multiplications that need to be implemented in separate digital computing units. Therefore, power-hungry analogue-to-digital conversion is needed at the crossbar outputs, which limits energy efficiency. Moreover, a fast and flexible communication scheme together with highly-efficient pipelining of the digital compute units and intermediate SRAM storage is primordial to ensure that they do not dominate the latency and power consumption. Finally, a user-friendly software-stack that tightly integrates such hardware with common machine learning frameworks (Pytorch/Tensorflow) is key for its widespread adoption within the community.

8.1.3.Potential Solutions

The advancement in the field requires a joint development of both material aspects of the memory, algorithms used to train the neural networks, efficient network mapping techniques over multiple crossbar arrays and careful design of the peripheral circuitry. In general, inference/training chips are required to have high-accuracy computation and high energy (operations per energy) and high area (operations per millimetre square) efficiency.

To improve the compute precision up to four or five-bit fixed-point arithmetic, it is essential to minimize the temporal conductance fluctuations (such as noise, conductance drift and temperature dependence). To achieve this, additional material research such as proper incorporation of dopants, memories composed of multi-layer materials stacks ⁴⁶¹, and exploiting material confinement to change device properties ³³⁰, is essential. To improve the compute density, besides scaling both the devices and the associated access transistors, high-density arrays need to be integrated at the back end of a CMOS wafer. To decrease the computational time required to convert the integrated charge/voltage after a Multiply-and-Accumulate (MAC) operation, column multiplexing should be avoided, using a per-column circuitry, shrinking the area to accommodate Analog-to-Digital Converters (ADCs). Exploration of ADC designs with low-number of bits enables efficient integration, while still keeping a reasonably high MAC accuracy ⁴⁵⁹. Another approach uses fully analog peripheral circuitry, which is more power efficient ⁴⁵⁴, posing however more stringent limitations on computation precision and available activation functions. As an example, ReLU has been demonstrated in the analog domain, while non-linear Sigmoid or Tanh generally require digital processing.

Even in the case of highly-efficient crossbar arrays, the general performance could still be poor due to Amdahl’s Law, since fast analog cores would generate large amounts of data that the neighbouring digital cores need to process ⁴⁶⁰. In other words, highly efficient analog circuitry would require highly efficient digital blocks closely located. Therefore, to improve the hybrid analog/digital chip, spatial network mapping, analog/digital cores spatial location and data communication need to be codesigned, leading to a general trade-off between chip reconfigurability (ability to map any type of network) and performance (ability to efficiently process specific networks) ⁴⁶⁰.

8.1.4.Conclusion

Computing in memory has greatly improved in recent years thanks to several on-chip demonstrations. However, challenges need to be overcome to get to product level, such as 4-to-5 bit computing precision together with performances in the range of 100/200 TOPS/W. Achieving such specifications is critical to provide a competitive advantage over existing purely digital processing cores. This will be obtained by optimizing the full computing stack: devices will require large uniformity, reasonably high resistance levels  and overall low noise behavior. Compact peripheral circuitry will be essential to reduce the chip area, and careful spatial mapping of analog and digital dedicated cores will need to improve the performance on a variety of neural networks.

9. Epilogue and Concluding Remarks for the Roadmap

Ilia Valov, Rainer Waser, Adnan Mehonic

After decades of reliance on transistor-based electronics, we are now delving into an era where the exploration of innovative nanoelectronic technologies based on functional materials is more crucial than ever. For instance, nanoscale memristive devices have emerged as key components for future nanoelectronics. Their straightforward stack structure, diverse functionalities, and specific benefits such as scalability, broad temperature stability and operation range, and resilience to high-energy particles and electromagnetic interference make them indispensable for numerous applications. Furthermore, memristors and similar novel technologies now serve as foundational units for the next generation of brain-inspired computing architectures. From their introduction in the ’60s of the last century as resistive switching memories, through the relation to Chua’s memristor till nowadays using them as artificial neurons and synapses, memristive devices have passed decades of intensive research with respect to both fundamentals and applications by academia and industry.

Examining the foundational aspects of materials science is pivotal for developing new nanoelectronic technologies. Using memristive technology as an example: the appeal of memristive devices stems from the multitude of benefits they provide, which are influenced and modulated by the materials and processes that dictate their behaviour and functionalities. Here especially the relation between materials, materials properties, physicochemical processes, and functionalities should be highlighted. A huge spectrum of materials has been used for switching films – 1D (single molecules or molecular clusters), 2D (Graphene, hBN, MoS2, MoSe2 etc.) and 3D, including inorganic, organic and biomaterials.  Several physical phenomena were reported to lead to memristive behaviour - phase change, redox reactions and ionic transport, electronic effects, van der Waals forces, and magnetic and magneto-resistance changes, all covered in this roadmap.

All different phenomena, of course, depend strongly on the used materials. The main challenge appears to properly select a combination of appropriate materials and their dimensions (i.e. thickness and lateral scale). Apparently, a simple two-electrode cell is composed of a switching layer or in many cases more switching layers, two electrodes and capping film(s). The devices are exposed to extreme operating conditions such as current densities of up to 1 million Amperes per square centimetre and electric fields of 10⁸ Volts per meter. Due to these extreme conditions and the nano-dimensions of the films their thermodynamic and kinetic behaviour typically deviate strongly from their macroscopic counterparts going much beyond the frame of the classical knowledge in terms of chemical and mechanical stability, point defect chemistry and transport properties. The nano-size effects are especially prominent considering the switching films. These are typically materials, that in a macroscopic sense are insulators, some used as well as high-k dielectrics, but turn into mixed solid electrolytes at thicknesses below ~ 50 nm. Furthermore, due to enhanced surface energy excess, the layers react chemically within the stack even if the change in the Gibbs energy for the reaction is positive. As a result, intermediate films can be formed, with similar thicknesses as the switching layers, that can either inhibit or enhance charge and mass transport through interfaces, strongly influencing all related processes. Other effects caused by impurities, absorption of moisture and/or oxygen, local changes of concentrations and nanocavities are known to additionally complicate the control over the device’s behaviour and characteristics. This poses a challenge, yet also presents an opportunity to uncover unique device physics and explore alternative device functionalities (e.g. neuromorphic functionalities).

The way to keep control over the devices and to further expand the horizon of functionalities is by using a material science-based approach where materials properties and processes are studied in very detail and this knowledge should be applied in the design of the devices.

The “Roadmap to Neuromorphic Computing with Emerging Technologies” roadmap seeks to tackle these issues and present a contemporary perspective on the intersections between current materials science, electronic engineering, and system design. Its ultimate goal is to delve into alternative computing models, particularly brain-inspired (neuromorphic) computing. Eminent groups and experts span a wide spectrum of relevant subjects and technologies, fostering a platform for idea exchange. Additionally, they highlight next-generation neuromorphic hardware, emphasizing the foundational role of functional materials and innovative device technologies. Although materials have been the main focus of the roadmap, our aim was to provide a more holistic overview and highlight a range of emerging and highly active research areas. As such, there are many details and specific material considerations not covered here that we strongly recommend the authors explore in the extensive background literature, much of which is published in special issues ¹³, ⁴⁶², ⁴⁶³, ⁴⁶⁴or excellent reviews ¹², ⁸. Likewise, in order to keep the format of the roadmap relatively compact, we have not elaborated on a number of equally valid and highly promising approaches in the context of neuromorphic technology development. Some notable examples include  adaptive matters and computation based on disorder  ²⁸⁷, ⁴⁶⁵, systems based on organic  perovskites ^466,467,  ionic-liquid based devices ^468,469, and molecular devices ^470,471,63. Of course, the list is not exhaustive, and other approaches, physical systems, and technologies are emerging.

Likewise, in order to keep the format of the roadmap in relatively conpact form, we have not elaborated on a number of equally valid and highly promising approahces in context of neuromorphic technology development.

Finally, recommending the most promising material systems is a complex task. In general, RRAM technology might have an advantage due to its simplicity and CMOS-friendliness. However, recent developments in HfO2-based FeRAMs represent a highly active area of research. MRAM is likely the most mature technology, with already available products, while PCM has seen significant interest recently. 2D materials are expected to integrate with all these technologies, providing further device improvements. It is important to keep in mind that the requirements might be dramatically different depending on targeted applications. For embedded systems, ease of integration and full CMOS compatibility are likely the most important factors. Conventional NVM devices need to outperform Flash, while in the context of computing, higher endurance will likely be needed. Additionally, the requirements for less conventional analog or neuromorphic functionalities are somewhat less defined but equally relevant and in development. One should also bear in mind the gap that exists between academic research on proof-of-concept demonstrators and industrial R&D. Industrial R&D must consider not only technical factors but also economic feasibility ⁴⁷².

Neuromorphic technologies are undoubtedly poised to be strong contenders for the future of computing, whether based on conventional digital, analog, or conceptually different computing and signal processing paradigms.

References

1 A. Mehonic and A.J. Kenyon, Nature 604, 255 (2022).

2 T. Gokmen and Y. Vlasov, Front Neurosci 10, 333 (2016).

3 M.J. Marinella, S. Agarwal, A. Hsia, I. Richter, R. Jacobs-Gedrim, J. Niroula, S.J. Plimpton, E. Ipek, and C.D. James, IEEE Journal on Emerging and Selected Topics in Circuits and Systems 8, 86 (2018).

4 H.-Y. Chang, P. Narayanan, S.C. Lewis, N.C.P. Farinha, K. Hosokawa, C. Mackin, H. Tsai, S. Ambrogio, A. Chen, and G.W. Burr, IBM Journal of Research and Development 63, 8 (2019).

5 C. Mead, Nature Electronics 3, 434 (2020).

6 A. Mehonic and A.J. Kenyon, Frontiers in Neuroscience 10, (2016).

7 S. Kumar, R.S. Williams, and Z. Wang, Nature 585, (2020).

8 P. Mannocci, M. Farronato, N. Lepri, L. Cattaneo, A. Glukhov, Z. Sun, and D. Ielmini, APL Machine Learning 1, (2023).

9 D.V. Christensen, R. Dittmann, B. Linares-Barranco, A. Sebastian, M.L. Gallo, A. Redaelli, S. Slesazeck, T. Mikolajick, S. Spiga, S. Menzel, I. Valov, G. Milano, C. Ricciardi, S.-J. Liang, F. Miao, M. Lanza, T.J. Quill, S.T. Keene, A. Salleo, J. Grollier, D. Marković, A. Mizrahi, P. Yao, J.J. Yang, G. Indiveri, J.P. Strachan, S. Datta, E. Vianello, A. Valentian, J. Feldmann, X. Li, W.H.P. Pernice, H. Bhaskaran, S. Furber, E. Neftci, F. Scherr, W. Maass, S. Ramaswamy, J. Tapson, P. Panda, Y. Kim, G. Tanaka, S. Thorpe, C. Bartolozzi, T.A. Cleland, C. Posch, S.C. Liu, G. Panuccio, M. Mahmud, A.N. Mazumder, M. Hosseini, T. Mohsenin, E. Donati, S. Tolu, R. Galeazzi, M.E. Christensen, S. Holm, D. Ielmini, and N. Pryds, Neuromorphic Computing and Engineering 2, 022501 (2022).

10 D. Ielmini, Z. Wang, and Y. Liu, APL Materials 9, 050702 (2021).

11 G. Li, L. Deng, H. Tang, G. Pan, Y. Tian, K. Roy, and W. Maass, (2023).

12 M.-K. Kim, Y. Park, I.-J. Kim, and J.-S. Lee, IScience 23, (2020).

13 I.K. Schuller, A. Frano, R.C. Dynes, A. Hoffmann, B. Noheda, C. Schuman, A. Sebastian, and J. Shen, Applied Physics Letters 120, (2022).

14 F. Torres, A.C. Basaran, and I.K. Schuller, Advanced Materials 35, (2023).

15 A.W. Burks, H.H. Goldstine, and J. Neumann, in The Origins of Digital Computers (Springer Berlin Heidelberg, 1982), pp. 399–413.

16 R.R. Schaller, IEEE Spectrum 34, 52 (1997).

17 M.J. Flynn, Proceedings of the IEEE 54, 1901 (1966).

18 R.H. Dennard, F.H. Gaensslen, H.-N. Yu, V.L. Rideout, E. Bassous, and A.R. LeBlanc, IEEE Journal of Solid-State Circuits 9, 256 (1974).

19 H. Esmaeilzadeh, E. Blem, R.S. Amant, K. Sankaralingam, and D. Burger, in Proceedings of the 38th Annual International Symposium on Computer Architecture (ACM, 2011).

20 H.-S.P. Wong, H.-Y. Lee, S. Yu, Y.-S. Chen, Y. Wu, P.-S. Chen, B. Lee, F.T. Chen, and M.-J. Tsai, Proceedings of the IEEE 100, 1951 (2012).

21 M. Horowitz, in 2014 IEEE International Solid-State Circuits Conference Digest of Technical Papers (ISSCC) (IEEE, 2014).

A. Pedram, S. Richardson, M. Horowitz, S. Galal, and S. Kvatinsky, IEEE Design & Test 34, 39 (2017).

23 A. Boroumand, S. Ghose, Y. Kim, R. Ausavarungnirun, E. Shiu, R. Thakur, D. Kim, A. Kuusela, A. Knies, P. Ranganathan, and O. Mutlu, ACM SIGPLAN Notices 53, 316 (2018).

24 A. Boroumand, S. Ghose, B. Akin, R. Narayanaswami, G.F. Oliveira, X. Ma, E. Shiu, and O. Mutlu, in 2021 30th International Conference on Parallel Architectures and Compilation Techniques (PACT) (IEEE, 2021).

25 S. Palacharla and R.E. Kessler, in Proceedings of 21 International Symposium on Computer Architecture (IEEE Comput. Soc. Press, n.d.).

26 R. Bera, K. Kanellopoulos, S. Balachandran, D. Novo, A. Olgun, M. Sadrosadat, and O. Mutlu, in 2022 55th IEEE/ACM International Symposium on Microarchitecture (MICRO) (IEEE, 2022).

27 S. Srinath, O. Mutlu, H. Kim, and Y.N. Patt, in 2007 IEEE 13th International Symposium on High Performance Computer Architecture (IEEE, 2007).

28Supercomputing Frontiers and Innovations 1, (2014).

29 O. Mutlu, S. Ghose, J. Gómez-Luna, and R. Ausavarungnirun, in Emerging Computing: From Devices to Systems (Springer Nature Singapore, 2022), pp. 171–243.

30 F. Devaux, in 2019 IEEE Hot Chips 31 Symposium (HCS) (IEEE, 2019).

31 Y.-C. Kwon, S.H. Lee, J. Lee, S.-H. Kwon, J.M. Ryu, J.-P. Son, O. Seongil, H.-S. Yu, H. Lee, S.Y. Kim, Y. Cho, J.G. Kim, J. Choi, H.-S. Shin, J. Kim, B.S. Phuah, H.M. Kim, M.J. Song, A. Choi, D. Kim, S.Y. Kim, E.-B. Kim, D. Wang, S. Kang, Y. Ro, S. Seo, J.H. Song, J. Youn, K. Sohn, and N.S. Kim, in 2021 IEEE International Solid- State Circuits Conference (ISSCC) (IEEE, 2021).

32 J. Borghetti, G.S. Snider, P.J. Kuekes, J.J. Yang, D.R. Stewart, and R.S. Williams, Nature 464, 873 (2010).

33 S. Kvatinsky, D. Belousov, S. Liman, G. Satat, N. Wald, E.G. Friedman, A. Kolodny, and U.C. Weiser, IEEE Transactions on Circuits and Systems II: Express Briefs 61, 895 (2014).

34 V. Seshadri, D. Lee, T. Mullins, H. Hassan, A. Boroumand, J. Kim, M.A. Kozuch, O. Mutlu, P.B. Gibbons, and T.C. Mowry, in Proceedings of the 50th Annual IEEE/ACM International Symposium on Microarchitecture (ACM, 2017).

35 A. Shafiee, A. Nag, N. Muralimanohar, R. Balasubramonian, J.P. Strachan, M. Hu, R.S. Williams, and V. Srikumar, ACM SIGARCH Computer Architecture News 44, 14 (2016).

36 N. Hajinazar, G.F. Oliveira, S. Gregorio, J.D. Ferreira, N.M. Ghiasi, M. Patel, M. Alser, S. Ghose, J. Gómez-Luna, and O. Mutlu, in Proceedings of the 26th ACM International Conference on Architectural Support for Programming Languages and Operating Systems (ACM, 2021).

37 M. Prezioso, F. Merrikh-Bayat, B.D. Hoskins, G.C. Adam, K.K. Likharev, and D.B. Strukov, Nature 521, 61 (2015).

38 C. Posch, T. Serrano-Gotarredona, B. Linares-Barranco, and T. Delbruck, Proceedings of the IEEE 102, 1470 (2014).

39 T. Serrano-Gotarredona and B. Linares-Barranco, IEEE Journal of Solid-State Circuits 48, 827 (2013).

40 P. Lichtsteiner, C. Posch, and T. Delbruck, IEEE Journal of Solid-State Circuits 43, 566 (2008).

41 D.P. Moeys, C. Li, J.N.P. Martel, S. Bamford, L. Longinotti, V. Motsnyi, D.S.S. Bello, and T. Delbruck, in 2017 IEEE International Symposium on Circuits and Systems (ISCAS) (IEEE, 2017).

42 T. Finateu, A. Niwa, D. Matolin, K. Tsuchimoto, A. Mascheroni, E. Reynaud, P. Mostafalu, F. Brady, L. Chotard, F. LeGoff, H. Takahashi, H. Wakabayashi, Y. Oike, and C. Posch, in 2020 IEEE International Solid- State Circuits Conference - (ISSCC) (IEEE, 2020).

43 A. van Schaik and S.-C. Liu, in 2005 IEEE International Symposium on Circuits and Systems (IEEE, n.d.).

44 S.-W. Chiu and K.-T. Tang, Sensors 13, 14214 (2013).

45 S. Caviglia, M. Valle, and C. Bartolozzi, in 2014 IEEE International Symposium on Circuits and Systems (ISCAS) (IEEE, 2014).

46 J.A. Perez-Carrasco, B. Zhao, C. Serrano, B. Acha, T. Serrano-Gotarredona, S. Chen, and B. Linares-Barranco, IEEE Transactions on Pattern Analysis and Machine Intelligence 35, 2706 (2013).

47 E. Painkras, L.A. Plana, J. Garside, S. Temple, F. Galluppi, C. Patterson, D.R. Lester, A.D. Brown, and S.B. Furber, IEEE Journal of Solid-State Circuits 48, 1943 (2013).

48 H. Markram, Scientific American 306, 50 (2012).

49 J.S. Sebastian Millner Andreas Hartel and K. Meier, .

50 P.A. Merolla, J.V. Arthur, R. Alvarez-Icaza, A.S. Cassidy, J. Sawada, F. Akopyan, B.L. Jackson, N. Imam, C. Guo, Y. Nakamura, B. Brezzo, I. Vo, S.K. Esser, R. Appuswamy, B. Taba, A. Amir, M.D. Flickner, W.P. Risk, R. Manohar, and D.S. Modha, Science 345, 668 (2014).

51 M. Davies, N. Srinivasa, T.-H. Lin, G. Chinya, Y. Cao, S.H. Choday, G. Dimou, P. Joshi, N. Imam, S. Jain, Y. Liao, C.-K. Lin, A. Lines, R. Liu, D. Mathaikutty, S. McCoy, A. Paul, J. Tse, G. Venkataramanan, Y.-H. Weng, A. Wild, Y. Yang, and H. Wang, IEEE Micro 38, 82 (2018).

52 https://www.intel.com/content/www/us/en/newsroom/news/intel-scales-neuromorphic-research-system-100-million-neurons.html.

53 Y. LeCun, Y. Bengio, and G. Hinton, Nature 521, 436 (2015).

54 A. Tavanaei, M. Ghodrati, S.R. Kheradpisheh, T. Masquelier, and A. Maida, Neural Networks 111, 47 (2019).

55 L.A. Camunas-Mesa, E. Vianello, C. Reita, T. Serrano-Gotarredona, and B. Linares-Barranco, IEEE Journal on Emerging and Selected Topics in Circuits and Systems 12, 898 (2022).

56 W. Wan, R. Kubendran, C. Schaefer, S.B. Eryilmaz, W. Zhang, D. Wu, S. Deiss, P. Raina, H. Qian, B. Gao, S. Joshi, H. Wu, H.-S.P. Wong, and G. Cauwenberghs, Nature 608, 504 (2022).

57 B. Linares-Barranco, Nature Electronics 1, 100 (2018).

58 S. Furber, Journal of Neural Engineering 13, 051001 (2016).

59 H. Hendy and C. Merkel, Journal of Electronic Imaging 31, (2022).

60 A. Sebastian, M.L. Gallo, R. Khaddam-Aljameh, and E. Eleftheriou, Nature Nanotechnology 15, 529 (2020).

61 S. Brivio, S. Spiga, and D. Ielmini, Neuromorphic Computing and Engineering 2, 042001 (2022).

62 G.W. Burr, R.M. Shelby, A. Sebastian, S. Kim, S. Kim, S. Sidler, K. Virwani, M. Ishii, P. Narayanan, A. Fumarola, L.L. Sanches, I. Boybat, M. Le Gallo, K. Moon, J. Woo, H. Hwang, and Y. Leblebici, Advances in Physics: X 2, (2016).

63 V.K. Sangwan and M.C. Hersam, Nature Nanotechnology 15, (2020).

64 Z. Wang, H. Wu, G.W. Burr, C.S. Hwang, K.L. Wang, Q. Xia, and J.J. Yang, Nature Reviews Materials 5, 173 (2020).

65 J. Grollier, D. Querlioz, K.Y. Camsari, K. Everschor-Sitte, S. Fukami, and M.D. Stiles, Nature Electronics 3, 360 (2020).

66 G. Milano, G. Pedretti, K. Montano, S. Ricci, S. Hashemkhani, L. Boarino, D. Ielmini, and C. Ricciardi, Nat Mater 21, 195 (2022).

67 G.W. Burr, R.M. Shelby, A. Sebastian, S. Kim, S. Kim, S. Sidler, K. Virwani, M. Ishii, P. Narayanan, A. Fumarola, and others, Advances in Physics: X 2, 89 (2017).

68 A. Sebastian, M. Le Gallo, R. Khaddam-Aljameh, and E. Eleftheriou, Nature Nanotechnology 15, 529 (2020).

69 E. Covi, H. Mulaosmanovic, B. Max, S. Slesazeck, and T. Mikolajick, Neuromorphic Computing and Engineering 2, 012002 (2022).

70 A.A. Talin, Y. Li, D.A. Robinson, E.J. Fuller, and S. Kumar, Advanced Materials 2204771 (2022).

71 Y. van de Burgt, A. Melianas, S.T. Keene, G. Malliaras, and A. Salleo, Nature Electronics 1, 386 (2018).

72 R. Wang, J.-Q. Yang, J.-Y. Mao, Z.-P. Wang, S. Wu, M. Zhou, T. Chen, Y. Zhou, and S.-T. Han, Advanced Intelligent Systems 2, 2000055 (2020).

73 K. Zhu, S. Pazos, F. Aguirre, Y. Shen, Y. Yuan, W. Zheng, O. Alharbi, M.A. Villena, B. Fang, X. Li, A. Milozzi, M. Farronato, M. Muñoz-Rojo, T. Wang, R. Li, H. Fariborzi, J.B. Roldan, G. Benstetter, X. Zhang, H. Alshareef, T. Grasser, H. Wu, D. Ielmini, and M. Lanza, Nature (2023).

74 B.J. Shastri, A.N. Tait, T.F. de Lima, W.H.P. Pernice, H. Bhaskaran, C.D. Wright, and P.R. Prucnal, Nature Photonics 15, 102 (2021).

75 W. Zhou, N. Farmakidis, J. Feldmann, X. Li, J. Tan, Y. He, C.D. Wright, W.H.P. Pernice, and H. Bhaskaran, MRS Bulletin 47, 502 (2022).

76 S. Jung, H. Lee, S. Myung, H. Kim, S.K. Yoon, S.-W. Kwon, Y. Ju, M. Kim, W. Yi, S. Han, B. Kwon, B. Seo, K. Lee, G.-H. Koh, K. Lee, Y. Song, C. Choi, D. Ham, and S.J. Kim, Nature 601, 211 (2022).

77 M. Payvand, F. Moro, K. Nomura, T. Dalgaty, E. Vianello, Y. Nishi, and G. Indiveri, Nature Communications 13, (2022).

78 F. Moro, E. Hardy, B. Fain, T. Dalgaty, P. Clémençon, A.D. Prà, E. Esmanhotto, N. Castellani, F. Blard, F. Gardien, T. Mesquida, F. Rummens, D. Esseni, J. Casas, G. Indiveri, M. Payvand, and E. Vianello, Nature Communications 13, (2022).

79 S. Brivio, D. Conti, M.V. Nair, J. Frascaroli, E. Covi, C. Ricciardi, G. Indiveri, and S. Spiga, Nanotechnology 30, 015102 (2018).

80 Y. Demirag, F. Moro, T. Dalgaty, G. Navarro, C. Frenkel, G. Indiveri, E. Vianello, and M. Payvand, in 2021 IEEE International Symposium on Circuits and Systems (ISCAS) (IEEE, 2021).

81 L.G. Wright, T. Onodera, M.M. Stein, T. Wang, D.T. Schachter, Z. Hu, and P.L. McMahon, Nature 601, 549 (2022).

82 G. Tanaka, T. Yamane, J.B. Héroux, R. Nakane, N. Kanazawa, S. Takeda, H. Numata, D. Nakano, and A. Hirose, Neural Networks 115, 100 (2019).

83 S. Ghosh, K. Nakajima, T. Krisnanda, K. Fujii, and T.C.H. Liew, Advanced Quantum Technologies 4, 2100053 (2021).

84 G. Csaba and W. Porod, Applied Physics Reviews 7, 011302 (2020).

85 K. Yamazaki, V.-K. Vo-Ho, D. Bulsara, and N. Le, Brain Sciences 12, 863 (2022).

86 J. Zhu, T. Zhang, Y. Yang, and R. Huang, Applied Physics Reviews 7, 011312 (2020).

87 D. Ielmini and S. Ambrogio, Nanotechnology 31, 092001 (2019).

88 G. Zhou, Z. Wang, B. Sun, F. Zhou, L. Sun, H. Zhao, X. Hu, X. Peng, J. Yan, H. Wang, W. Wang, J. Li, B. Yan, D. Kuang, Y. Wang, L. Wang, and S. Duan, Advanced Electronic Materials 8, 2101127 (2022).

89 M.D. Pickett, G. Medeiros-Ribeiro, and R.S. Williams, Nature Materials 12, 114 (2012).

90 Z. Wang, S. Joshi, S.E. Savel’ev, H. Jiang, R. Midya, P. Lin, M. Hu, N. Ge, J.P. Strachan, Z. Li, Q. Wu, M. Barnell, G.-L. Li, H.L. Xin, R.S. Williams, Q. Xia, and J.J. Yang, Nature Materials 16, 101 (2016).

91 H. Jiang, D. Belkin, S.E. Savel’ev, S. Lin, Z. Wang, Y. Li, S. Joshi, R. Midya, C. Li, M. Rao, M. Barnell, Q. Wu, J.J. Yang, and Q. Xia, Nature Communications 8, (2017).

Y. Ushakov, A. Balanov, and S. Savel’ev, Chaos, Solitons & Fractals 145, 110803 (2021).

93 Z. Wang, S. Joshi, S. Savel’ev, W. Song, R. Midya, Y. Li, M. Rao, P. Yan, S. Asapu, Y. Zhuo, H. Jiang, P. Lin, C. Li, J.H. Yoon, N.K. Upadhyay, J. Zhang, M. Hu, J.P. Strachan, M. Barnell, Q. Wu, H. Wu, R.S. Williams, Q. Xia, and J.J. Yang, Nature Electronics 1, 137 (2018).

94 A. Mehonic, A. Sebastian, B. Rajendran, O. Simeone, E. Vasilaki, and A.J. Kenyon, Advanced Intelligent Systems 2, 2000085 (2020).

95 D. Kim, B. Jeon, Y. Lee, D. Kim, Y. Cho, and S. Kim, Applied Physics Letters 121, 010501 (2022).

Y. Ushakov, A. Akther, P. Borisov, D. Pattnaik, S. Savel’ev, and A.G. Balanov, Chaos, Solitons & Fractals 149, 110997 (2021).

A.M. Wojtusiak, A.G. Balanov, and S.E. Savel’ev, Chaos, Solitons & Fractals 142, 110383 (2021).

98 D. Sussillo, Current Opinion in Neurobiology 25, 156 (2014).

99 S. Gepshtein, A.S. Pawar, S. Kwon, S. Savel’ev, and T.D. Albright, Science Advances 8, (2022).

100 J.D. Monaco, K. Rajan, and G.M. Hwang, ArXiv abs/2105.07284, (2021).

101 S. Kumar, X. Wang, J.P. Strachan, Y. Yang, and W.D. Lu, Nature Reviews Materials 7, (2022).

102 J. Feldmann, N. Youngblood, M. Karpov, H. Gehring, X. Li, M. Stappers, M.L. Gallo, X. Fu, A. Lukashchuk, A.S. Raja, J. Liu, C.D. Wright, A. Sebastian, T.J. Kippenberg, W.H.P. Pernice, and H. Bhaskaran, Nature 589, 52 (2021).

103 X. Wang, Y. Chen, H. Xi, H. Li, and D. Dimitrov, IEEE Electron Device Letters 30, 294 (2009).

104 M. Mansueto, A. Chavent, S. Auffret, I. Joumard, L. Vila, R.C. Sousa, L.D. Buda-Prejbeanu, I.L. Prejbeanu, and B. Dieny, Nanoscale 13, 11488 (2021).

105 A. Hirohata, K. Yamada, Y. Nakatani, I.-L. Prejbeanu, B. Diény, P. Pirro, and B. Hillebrands, Journal of Magnetism and Magnetic Materials 509, (2020).

106 A.H. Jaafar, R.J. Gray, E. Verrelli, M. O'Neill, S.M. Kelly, and N.T. Kemp, Nanoscale 9, 17091 (2017).

107 L.E. Srouji, A. Krishnan, R. Ravichandran, Y. Lee, M. On, X. Xiao, and S.J.B. Yoo, APL Photonics 7, 051101 (2022).

108 J.C. Gartside, K.D. Stenning, A. Vanstone, H.H. Holder, D.M. Arroo, T. Dion, F. Caravelli, H. Kurebayashi, and W.R. Branford, Nature Nanotechnology 17, 460 (2022).

109 D.D. Yaremkevich, A.V. Scherbakov, L. De Clerk, S.M. Kukhtaruk, A. Nadzeyka, R. Campion, A.W. Rushforth, S. Savel’ev, A.G. Balanov, and M. Bayer, Nature Communications 14, (2023).

110 K. Vandoorne, P. Mechet, T.V. Vaerenbergh, M. Fiers, G. Morthier, D. Verstraeten, B. Schrauwen, J. Dambre, and P. Bienstman, Nature Communications 5, (2014).

111 D.J. Gauthier, E. Bollt, A. Griffith, and W.A.S. Barbosa, Nature Communications 12, (2021).

112 S.-in Yi, J.D. Kendall, R.S. Williams, and S. Kumar, Nature Electronics (2022).

113 A. Acín, I. Bloch, H. Buhrman, T. Calarco, C. Eichler, J. Eisert, D. Esteve, N. Gisin, S.J. Glaser, F. Jelezko, S. Kuhr, M. Lewenstein, M.F. Riedel, P.O. Schmidt, R. Thew, A. Wallraff, I. Walmsley, and F.K. Wilhelm, New Journal of Physics 20, 080201 (2018).

114 V. Dunjko and H.J. Briegel, Reports on Progress in Physics 81, 074001 (2018).

115 P. Mujal, R. Martínez-Peña, J. Nokkala, J. García-Beni, G.L. Giorgi, M.C. Soriano, and R. Zambrini, Advanced Quantum Technologies 4, 2100027 (2021).

116 P. Pfeiffer, I.L. Egusquiza, M.D. Ventra, M. Sanz, and E. Solano, Scientific Reports 6, (2016).

117 J. Salmilehto, F. Deppe, M.D. Ventra, M. Sanz, and E. Solano, Scientific Reports 7, (2017).

118 M. Spagnolo, J. Morris, S. Piacentini, M. Antesberger, F. Massa, A. Crespi, F. Ceccarelli, R. Osellame, and P. Walther, Nature Photonics 16, 318 (2022).

119 K. Friston, Nature Reviews Neuroscience 11, 127 (2010).

120 N.B. Janson and C.J. Marsden, Scientific Reports 7, (2017).

121 H.-T. Zhang, T.J. Park, A.N.M.N. Islam, D.S.J. Tran, S. Manna, Q. Wang, S. Mondal, H. Yu, S. Banik, S. Cheng, H. Zhou, S. Gamage, S. Mahapatra, Y. Zhu, Y. Abate, N. Jiang, S.K.R.S. Sankaranarayanan, A. Sengupta, C. Teuscher, and S. Ramanathan, Science 375, 533 (2022).

122 M.J. Colbrook, V. Antun, and A.C. Hansen, Proceedings of the National Academy of Sciences 119, (2022).

123 T. Song, W. Rim, S. Park, Y. Kim, J. Jung, G. Yang, S. Baek, J. Choi, B. Kwon, Y. Lee, S. Kim, G. Kim, H.-S. Won, J.-H. Ku, S.S. Paak, E.S. Jung, S.S. Park, and K. Kim, in 2016 IEEE International Solid-State Circuits Conference (ISSCC) (IEEE, 2016).

124 T. Song, J. Jung, W. Rim, H. Kim, Y. Kim, C. Park, J. Do, S. Park, S. Cho, H. Jung, B. Kwon, H.-S. Choi, J.S. Choi, and J.S. Yoon, in 2018 IEEE International Solid - State Circuits Conference - (ISSCC) (IEEE, 2018).

125 G. Yeap, X. Chen, B.R. Yang, C.P. Lin, F.C. Yang, Y.K. Leung, D.W. Lin, C.P. Chen, K.F. Yu, D.H. Chen, C.Y. Chang, S.S. Lin, H.K. Chen, P. Hung, C.S. Hou, Y.K. Cheng, J. Chang, L. Yuan, C.K. Lin, C.C. Chen, Y.C. Yeo, M.H. Tsai, Y.M. Chen, H.T. Lin, C.O. Chui, K.B. Huang, W. Chang, H.J. Lin, K.W. Chen, R. Chen, S.H. Sun, Q. Fu, H.T. Yang, H.L. Shang, H.T. Chiang, C.C. Yeh, T.L. Lee, C.H. Wang, S.L. Shue, C.W. Wu, R. Lu, W.R. Lin, J. Wu, F. Lai, P.W. Wang, Y.H. Wu, B.Z. Tien, Y.C. Huang, L.C. Lu, J. He, Y. Ku, J. Lin, M. Cao, T.S. Chang, S.M. Jang, H.C. Lin, Y.C. Peng, J.Y. Sheu, and M. Wang, in 2019 IEEE International Electron Devices Meeting (IEDM) (IEEE, 2019).

126 Y. Kim, C. Ong, A.M. Pillai, H. Jagadeesh, G. Baek, I. Rajwani, Z. Guo, and E. Karl, IEEE Journal of Solid-State Circuits 1 (2022).

127 F. Dong, X. Si, and M.-F. Chang, in 2022 IEEE 16th International Conference on Solid-State  Integrated Circuit Technology (ICSICT) (IEEE, 2022).

128 X. Si, M.-F. Chang, W.-S. Khwa, J.-J. Chen, J.-F. Li, X. Sun, R. Liu, S. Yu, H. Yamauchi, and Q. Li, IEEE Transactions on Circuits and Systems I: Regular Papers 66, 4172 (2019).

129 Y. Wang, S. Zhang, Y. Li, J. Chen, W. Zhao, and Y. Ha, IEEE Transactions on Very Large Scale Integration (VLSI) Systems 31, 684 (2023).

130 Q. Dong, S. Jeloka, M. Saligane, Y. Kim, M. Kawaminami, A. Harada, S. Miyoshi, M. Yasuda, D. Blaauw, and D. Sylvester, IEEE Journal of Solid-State Circuits 53, 1006 (2018).

131 B. Pan, G. Wang, H. Zhang, W. Kang, and W. Zhao, IEEE Transactions on Circuits and Systems II: Express Briefs 69, (2022).

132 J. Zhang, Z. Wang, and N. Verma, IEEE Journal of Solid-State Circuits 52, (2017).

133 A.S. Rekhi, B. Zimmer, N. Nedovic, N. Liu, R. Venkatesan, M. Wang, B. Khailany, W.J. Dally, and C.T. Gray, in Proceedings of the 56th Annual Design Automation Conference 2019 (ACM, 2019).

134 S.K. Gonugondla, C. Sakr, H. Dbouk, and N.R. Shanbhag, in Proceedings of the 39th International Conference on Computer-Aided Design (ACM, 2020).

135 B. Murmann, IEEE Transactions on Very Large Scale Integration (VLSI) Systems 29, 3 (2021).

136 S. Spetalnick and A. Raychowdhury, IEEE Transactions on Circuits and Systems I: Regular Papers 69, 1466 (2022).

J. Saikia, S. Yin, S.K. Cherupally, B. Zhang, J. Meng, M. Seok, and J.-S. Seo, in 2021 Design Automation & Test in Europe Conference & Exhibition (DATE) (IEEE, 2021).

138 Y.-D. Chih, P.-H. Lee, H. Fujiwara, Y.-C. Shih, C.-F. Lee, R. Naous, Y.-L. Chen, C.-P. Lo, C.-H. Lu, H. Mori, W.-C. Zhao, D. Sun, M.E. Sinangil, Y.-H. Chen, T.-L. Chou, K. Akarvardar, H.-J. Liao, Y. Wang, M.-F. Chang, and T.-Y.J. Chang, in 2021 IEEE International Solid- State Circuits Conference (ISSCC) (IEEE, 2021).

139 G. Desoli, N. Chawla, T. Boesch, M. Avodhyawasi, H. Rawat, H. Chawla, V.S. Abhijith, P. Zambotti, A. Sharma, C. Cappetta, M. Rossi, A. De Vita, and F. Girardi, in 2023 IEEE International Solid- State Circuits Conference (ISSCC) (IEEE, 2023).

140 J. Zhang, Z. Wang, and N. Verma, in 2016 IEEE Symposium on VLSI Circuits (VLSI-Circuits) (IEEE, 2016).

141 M. Kang, S.K. Gonugondla, A. Patil, and N.R. Shanbhag, IEEE Journal of Solid-State Circuits 53, 642 (2018).

142 H. Valavi, P.J. Ramadge, E. Nestler, and N. Verma, in 2018 IEEE Symposium on VLSI Circuits (IEEE, 2018).

143 H. Jia, H. Valavi, Y. Tang, J. Zhang, and N. Verma, in 2019 IEEE Hot Chips 31 Symposium (HCS) (IEEE, 2019).

144 H. Jia, M. Ozatay, Y. Tang, H. Valavi, R. Pathak, J. Lee, and N. Verma, in 2021 IEEE International Solid- State Circuits Conference (ISSCC) (IEEE, 2021).

145 K. Ueyoshi, I.A. Papistas, P. Houshmand, G.M. Sarda, V. Jain, M. Shi, Q. Zheng, S. Giraldo, P. Vrancx, J. Doevenspeck, D. Bhattacharjee, S. Cosemans, A. Mallik, P. Debacker, D. Verkest, and M. Verhelst, in 2022 IEEE International Solid- State Circuits Conference (ISSCC) (IEEE, 2022).

146 R. Bez, E. Camerlenghi, A. Modelli, and A. Visconti, Proc. IEEE 91, 489 (2003).

147 C.M. Compagnoni, A. Goda, A.S. Spinelli, P. Feeley, A.L. Lacaita, and A. Visconti, Proc. IEEE 105, 1609 (2017).

148 T. Pekny, L. Vu, J. Tsai, D. Srinivasan, E. Yu, J. Pabustan, J. Xu, S. Deshmukh, K.-F. Chan, M. Piccardi, K. Xu, G. Wang, K. Shakeri, V. Patel, T. Iwasaki, T. Wang, P. Musunuri, C. Gu, A. Mohammadzadeh, A. Ghalam, V. Moschiano, T. Vali, J. Park, J. Lee, and R. Ghodsi, IEEE Int. Solid-State Circuits Conf. (ISSCC) Dig. Tech. Papers 132 (2022).

149 F. Merrikh-Bayat, X. Guo, M. Klachko, M. Prezioso, K.K. Likharev, and D.B. Strukov, IEEE Trans. Neural Netw. Learn. Syst 29, 4782 (2018).

150 X. Guo, F. Merrikh-Bayat, M. Prezioso, Y. Chen, B. Nguyen, N. Do, and D.B. Strukov, in Proc. CICC (2017), pp. 1–4.

151 M. Bavandpour, S. Sahay, M.R. Mahmoodi, and D.B. Strukov, Neuromorph. Comput. Eng. 1, 014001 (2021).

152 P. Wang, F. Xu, B. Wang, B. Gao, H. Wu, H. Qian, and S. Yu, IEEE Trans. VLSI Syst. 27, 988 (2019).

153 W. Shim and S. Yu, IEEE Electron Device Lett. 42, 160 (2021).

154 S.-T. Lee, G. Yeom, H. Yoo, H.-S. Kim, S. Lim, J.-H. Bae, B.-G. Park, and J.-H. Lee, IEEE Trans. Electron Devices 68, 3365 (2021).

155 S.-T. Lee and J.-H. Lee, Front. Neurosci. 14, 14 (2020).

156 H.-T. Lue, P.-K. Hsu, M.-L. Wei, T.-H. Yeh, P.-Y. Du, W.-C. Chen, K.-C. Wang, and C.-Y. Lu, IEDM Tech. Dig. 915 (2019).

157 H.-N. Yoo, J.-W. Back, N.-H. Kim, D. Kwon, B.-G.- Park, and J.-H. Lee, Proc. IEEE Symp. VLSI Technol. Circuits 304 (2022).

158 G. Malavena, M. Filippi, A.S. Spinelli, and C.M. Compagnoni, IEEE Trans. Electron Devices 66, 4727 (2019).

159 G. Malavena, M. Filippi, A.S. Spinelli, and C.M. Compagnoni, IEEE Trans. Electron Devices 66, 4733 (2019).

160 C.M. Compagnoni and A.S. Spinelli, IEEE Transactions on Electron Devices 66, 4504 (2019).

161 G. Malavena, S. Petrò, A.S. Spinelli, and C.M. Compagnoni, Proc. ESSDERC 122 (2019).

162 D.-H. Kwon, K.M. Kim, J.H. Jang, J.M. Jeon, M.H. Lee, G.H. Kim, X.-S. Li, G.-S. Park, B. Lee, S. Han, M. Kim, and C.S. Hwang, Nature Nanotechnology 5, 148 (2010).

163 L. Song, Y. Zhuo, X. Qian, H. Li, and Y. Chen, in 2018 IEEE International Symposium on High Performance Computer Architecture (HPCA) (IEEE, 2018).

164 G. Dai, T. Huang, Y. Wang, H. Yang, and J. Wawrzynek, in Proceedings of the 24th Asia and South Pacific Design Automation Conference (ACM, 2019).

165 S. Wang, Y. Li, D. Wang, W. Zhang, X. Chen, D. Dong, S. Wang, X. Zhang, P. Lin, C. Gallicchio, X. Xu, Q. Liu, K.-T. Cheng, Z. Wang, D. Shang, and M. Liu, Nature Machine Intelligence 5, 104 (2023).

166 W. Zhang, S. Wang, Y. Li, X. Xu, D. Dong, N. Jiang, F. Wang, Z. Guo, R. Fang, C. Dou, K. Ni, Z. Wang, D. Shang, and M. Liu, in 2022 IEEE Symposium on VLSI Technology and Circuits (VLSI Technology and Circuits) (IEEE, 2022).

167 F. Alibart, E. Zamanidoost, and D.B. Strukov, Nature Communications 4, (2013).

168 F.M. Bayat, M. Prezioso, B. Chakrabarti, H. Nili, I. Kataeva, and D. Strukov, Nature Communications 9, (2018).

169 M. Hu, C.E. Graves, C. Li, Y. Li, N. Ge, E. Montgomery, N. Davila, H. Jiang, R.S. Williams, J.J. Yang, Q. Xia, and J.P. Strachan, Advanced Materials 30, 1705914 (2018).

170 Q. Duan, Z. Jing, X. Zou, Y. Wang, K. Yang, T. Zhang, S. Wu, R. Huang, and Y. Yang, Nature Communications 11, (2020).

171 M. Ishii, U. Shin, K. Hosokawa, M. BrightSky, W. Haensch, S. Kim, S. Lewis, A. Okazaki, J. Okazawa, M. Ito, M. Rasch, W. Kim, and A. Nomura, in 2019 IEEE International Electron Devices Meeting (IEDM) (IEEE, 2019).

172 S. Yu, Z. Li, P.-Y. Chen, H. Wu, B. Gao, D. Wang, W. Wu, and H. Qian, in 2016 IEEE International Electron Devices Meeting (IEDM) (IEEE, 2016).

173 B. Yan, Q. Yang, W.-H. Chen, K.-T. Chang, J.-W. Su, C.-H. Hsu, S.-H. Li, H.-Y. Lee, S.-S. Sheu, M.-S. Ho, Q. Wu, M.-F. Chang, Y. Chen, and H. Li, in 2019 Symposium on VLSI Technology (IEEE, 2019).

174 C.-X. Xue, T.-Y. Huang, J.-S. Liu, T.-W. Chang, H.-Y. Kao, J.-H. Wang, T.-W. Liu, S.-Y. Wei, S.-P. Huang, W.-C. Wei, Y.-R. Chen, T.-H. Hsu, Y.-K. Chen, Y.-C. Lo, T.-H. Wen, C.-C. Lo, R.-S. Liu, C.-C. Hsieh, K.-T. Tang, and M.-F. Chang, in 2020 IEEE International Solid- State Circuits Conference - (ISSCC) (IEEE, 2020).

175 Y. Lu, X. Li, L. Yan, T. Zhang, Y. Yang, Z. Song, and R. Huang, in 2020 IEEE International Electron Devices Meeting (IEDM) (IEEE, 2020).

176 V. Joshi, M.L. Gallo, S. Haefeli, I. Boybat, S.R. Nandakumar, C. Piveteau, M. Dazzi, B. Rajendran, A. Sebastian, and E. Eleftheriou, Nature Communications 11, (2020).

177 H. Li, W.-C. Chen, A. Levy, C.-H. Wang, H. Wang, P.-H. Chen, W. Wan, W.-S. Khwa, H. Chuang, Y.-D. Chih, M.-F. Chang, H.-S.P. Wong, and P. Raina, IEEE Transactions on Electron Devices 68, 6637 (2021).

178 P. Yao, H. Wu, B. Gao, S.B. Eryilmaz, X. Huang, W. Zhang, Q. Zhang, N. Deng, L. Shi, H.-S.P. Wong, and H. Qian, Nature Communications 8, (2017).

179 P. Yao, H. Wu, B. Gao, J. Tang, Q. Zhang, W. Zhang, J.J. Yang, and H. Qian, Nature 577, 641 (2020).

180 S. Ambrogio, P. Narayanan, H. Tsai, R.M. Shelby, I. Boybat, C. di Nolfo, S. Sidler, M. Giordano, M. Bodini, N.C.P. Farinha, B. Killeen, C. Cheng, Y. Jaoudi, and G.W. Burr, Nature 558, 60 (2018).

181 G. Karunaratne, M. Schmuck, M.L. Gallo, G. Cherubini, L. Benini, A. Sebastian, and A. Rahimi, Nature Communications 12, (2021).

182 P.M. Sheridan, F. Cai, C. Du, W. Ma, Z. Zhang, and W.D. Lu, Nature Nanotechnology 12, 784 (2017).

183 W. Wan, R. Kubendran, S.B. Eryilmaz, W. Zhang, Y. Liao, D. Wu, S. Deiss, B. Gao, P. Raina, S. Joshi, H. Wu, G. Cauwenberghs, and H.-S.P. Wong, in 2020 IEEE International Solid- State Circuits Conference - (ISSCC) (IEEE, 2020).

184 S. Woźniak, A. Pantazi, T. Bohnstingl, and E. Eleftheriou, Nature Machine Intelligence 2, 325 (2020).

185 C. Du, F. Cai, M.A. Zidan, W. Ma, S.H. Lee, and W.D. Lu, Nature Communications 8, (2017).

186 J. Moon, W. Ma, J.H. Shin, F. Cai, C. Du, S.H. Lee, and W.D. Lu, Nature Electronics 2, 480 (2019).

187 G. Milano, G. Pedretti, K. Montano, S. Ricci, S. Hashemkhani, L. Boarino, D. Ielmini, and C. Ricciardi, Nature Materials 21, 195 (2021).

188 H. Tsai, S. Ambrogio, C. Mackin, P. Narayanan, R.M. Shelby, K. Rocki, A. Chen, and G.W. Burr, in 2019 Symposium on VLSI Technology (IEEE, 2019).

189 G. Karunaratne, M.L. Gallo, G. Cherubini, L. Benini, A. Rahimi, and A. Sebastian, Nature Electronics 3, 327 (2020).

190 Z. Liu, J. Tang, B. Gao, P. Yao, X. Li, D. Liu, Y. Zhou, H. Qian, B. Hong, and H. Wu, Nature Communications 11, (2020).

191 Y. Zhong, J. Tang, X. Li, B. Gao, H. Qian, and H. Wu, Nature Communications 12, (2021).

192 Z. Sun, G. Pedretti, A. Bricalli, and D. Ielmini, Science Advances 6, (2020).

193 Y.J. Jeong, J. Lee, J. Moon, J.H. Shin, and W.D. Lu, Nano Letters 18, 4447 (2018).

194 S. Choi, J.H. Shin, J. Lee, P. Sheridan, and W.D. Lu, Nano Letters 17, 3113 (2017).

195 A. Sebastian, T. Tuma, N. Papandreou, M.L. Gallo, L. Kull, T. Parnell, and E. Eleftheriou, Nature Communications 8, (2017).

196 S. Choi, S.H. Tan, Z. Li, Y. Kim, C. Choi, P.-Y. Chen, H. Yeon, S. Yu, and J. Kim, Nature Materials 17, 335 (2018).

197 J. Yang, M. Rao, H. Tang, J.-B. Wu, W. Song, M. Zhang, W. Yin, Y. Zhuo, F. Kiani, B. Chen, H. Liu, H.-Y. Chen, R. Midya, F. Ye, H. Jiang, Z. Wang, M. Wu, M. Hu, H. Wang, Q. Xia, G. Ge, and J. Li, (2022).

198 F. Kiani, J. Yin, Z. Wang, J.J. Yang, and Q. Xia, Science Advances 7, (2021).

199 T. Dalgaty, N. Castellani, C. Turck, K.-E. Harabi, D. Querlioz, and E. Vianello, Nature Electronics 4, 151 (2021).

200 Z. Wang, C. Li, P. Lin, M. Rao, Y. Nie, W. Song, Q. Qiu, Y. Li, P. Yan, J.P. Strachan, N. Ge, N. McDonald, Q. Wu, M. Hu, H. Wu, R.S. Williams, Q. Xia, and J.J. Yang, Nature Machine Intelligence 1, 434 (2019).

201 M.L. Gallo, A. Sebastian, R. Mathis, M. Manica, H. Giefers, T. Tuma, C. Bekas, A. Curioni, and E. Eleftheriou, Nature Electronics 1, 246 (2018).

202 F. Cai, S. Kumar, T.V. Vaerenbergh, X. Sheng, R. Liu, C. Li, Z. Liu, M. Foltin, S. Yu, Q. Xia, J.J. Yang, R. Beausoleil, W.D. Lu, and J.P. Strachan, Nature Electronics 3, 409 (2020).

203 M.R. Mahmoodi, M. Prezioso, and D.B. Strukov, Nature Communications 10, (2019).

204 M. Le Gallo and A. Sebastian, Journal of Physics D: Applied Physics 53, 213002 (2020).

205 S.G. Sarwat, Materials Science and Technology 33, 1890 (2017).

206 A. Sebastian, M. Le Gallo, G.W. Burr, S. Kim, M. BrightSky, and E. Eleftheriou, Journal of Applied Physics 124, 111101 (2018).

207 S.R. Ovshinsky and B. Pashmakov, MRS Online Proceedings Library (OPL) 803, (2003).

208 M.L. Gallo, R. Khaddam-Aljameh, M. Stanisavljevic, A. Vasilopoulos, B. Kersting, M. Dazzi, G. Karunaratne, M. Braendli, A. Singh, S.M. Mueller, and others, Nature Electronics (2023).

209 S.R. Nandakumar, M. Le Gallo, C. Piveteau, V. Joshi, G. Mariani, I. Boybat, G. Karunaratne, R. Khaddam-Aljameh, U. Egger, A. Petropoulos, and others, Frontiers in Neuroscience 14, 406 (2020).

210 A. Sebastian, T. Tuma, N. Papandreou, M. Le Gallo, L. Kull, T. Parnell, and E. Eleftheriou, Nature Communications 8, 1115 (2017).

211 G. Karunaratne, M. Hersche, J. Langeneager, G. Cherubini, M. Le Gallo, U. Egger, K. Brew, S. Choi, I. Ok, C. Silvestre, and others, in ESSCIRC 2022-IEEE 48th European Solid State Circuits Conference (ESSCIRC) (IEEE, 2022), pp. 105–108.

212 S. Ambrogio, N. Ciocchini, M. Laudato, V. Milo, A. Pirovano, P. Fantini, and D. Ielmini, Frontiers in Neuroscience 10, 56 (2016).

213 T. Tuma, M. Le Gallo, A. Sebastian, and E. Eleftheriou, IEEE Electron Device Letters 37, 1238 (2016).

214 T. Tuma, A. Pantazi, M. Le Gallo, A. Sebastian, and E. Eleftheriou, Nature Nanotechnology 11, 693 (2016).

215 https://download.intel.com/newsroom/2021/archive/2015-07-28-news-releases-intel-and-micron-produce-breakthrough-memory-technology.pdf.

216 F. Arnaud, P. Ferreira, F. Piazza, A. Gandolfo, P. Zuliani, P. Mattavelli, E. Gomiero, G. Samanni, J. Jasse, C. Jahan, and others, in International Electron Devices Meeting (IEDM) (IEEE, 2020), pp. 24–2.

217 M. Lanza, A. Sebastian, W.D. Lu, M. Le Gallo, M.-F. Chang, D. Akinwande, F.M. Puglisi, H.N. Alshareef, M. Liu, and J.B. Roldan, Science 376, 9979 (2022).

218 S. Raoux, Annual Review of Materials Research 39, 25 (2009).

219 J. Liang, S. Yeh, S.S. Wong, and H.-S.P. Wong, 4th IEEE International Memory Workshop 1 (2012).

220 S. Yoo, H.D. Lee, S. Lee, H. Choi, and T. Kim, IEEE Transactions on Electron Devices 67, 1454 (2020).

221 S.H. Lee, M.S. Kim, G.S. Do, S.G. Kim, H.J. Lee, J.S. Sim, N.G. Park, S.B. Hong, Y.H. Jeon, K.S. Choi, and others, in 2010 Symposium on VLSI Technology (IEEE, 2010), pp. 199–200.

222 T. Mikolajick, M.H. Park, L. Begon-Lours, and S. Slesazeck, Adv Mater 2206042 (2022).

223 H.P. McAdams, R. Acklin, T. Blake, X.-H. Du, J. Eliason, J. Fong, W.F. Kraus, D. Liu, S. Madan, T. Moise, S. Natarajan, N. Qian, Y. Qiu, K.A. Remack, J. Rodriguez, J. Roscher, A. Seshadri, and S.R. Summerfelt, IEEE Journal of Solid-State Circuits 39, 667 (2004).

224 S. Beyer, S. Dunkel, M. Trentzsch, J. Muller, A. Hellmich, D. Utess, J. Paul, D. Kleimaier, J. Pellerin, S. Muller, J. Ocker, A. Benoist, H. Zhou, M. Mennenga, M. Schuster, F. Tassan, M. Noack, A. Pourkeramati, F. Muller, M. Lederer, T. Ali, R. Hoffmann, T. Kampfe, K. Seidel, H. Mulaosmanovic, E.T. Breyer, T. Mikolajick, and S. Slesazeck, in 2020 IEEE International Memory Workshop (IMW) (IEEE, 2020).

225 E.Y. Tsymbal and H. Kohlstedt, Science 313, 181 (2006).

226 K. Asadi, M. Li, P.W.M. Blom, M. Kemerink, and D.M. de Leeuw, Materials Today 14, 592 (2011).

227 D. Bondurant, Ferroelectrics 112, 273 (1990).

228 T.S. Böscke, J. Müller, D. Bräuhaus, U. Schröder, and U. Böttger, Applied Physics Letters 99, 102903 (2011).

229 H. Mulaosmanovic, P.D. Lomenzo, U. Schroeder, S. Slesazeck, T. Mikolajick, and B. Max, in 2021 IEEE International Reliability Physics Symposium (IRPS) (IEEE, 2021).

230 R. Materlik, C. Künneth, and A. Kersch, Journal of Applied Physics 117, 134109 (2015).

231 H. Mulaosmanovic, S. Dunkel, M. Trentzsch, S. Beyer, E.T. Breyer, T. Mikolajick, and S. Slesazeck, IEEE Transactions on Electron Devices 67, 5804 (2020).

232 M. Materano, P.D. Lomenzo, A. Kersch, M.H. Park, T. Mikolajick, and U. Schroeder, Inorganic Chemistry Frontiers 8, 2650 (2021).

233 D.R. Islamov, V.A. Gritsenko, T.V. Perevalov, V.A. Pustovarov, O.M. Orlov, A.G. Chernikova, A.M. Markeev, S. Slesazeck, U. Schroeder, T. Mikolajick, and G.Y. Krasnikov, Acta Materialia 166, 47 (2019).

234 L. Bégon-Lours, M. Halter, F.M. Puglisi, L. Benatti, D.F. Falcone, Y. Popoff, D.D. Pineda, M. Sousa, and B.J. Offrein, Advanced Electronic Materials 8, 2101395 (2022).

235 P. Nukala, M. Ahmadi, Y. Wei, S. de Graaf, E. Stylianidis, T. Chakrabortty, S. Matzen, H.W. Zandbergen, A. Björling, D. Mannix, D. Carbone, B. Kooi, and B. Noheda, Science 372, 630 (2021).

236 S.S. Cheema, D. Kwon, N. Shanker, R. dos Reis, S.-L. Hsu, J. Xiao, H. Zhang, R. Wagner, A. Datar, M.R. McCarter, C.R. Serrao, A.K. Yadav, G. Karbasian, C.-H. Hsu, A.J. Tan, L.-C. Wang, V. Thakare, X. Zhang, A. Mehta, E. Karapetrova, R.V. Chopdekar, P. Shafer, E. Arenholz, C. Hu, R. Proksch, R. Ramesh, J. Ciston, and S. Salahuddin, Nature 580, 478 (2020).

237 Y. Wei, G. Vats, and B. Noheda, Neuromorphic Computing and Engineering 2, 044007 (2022).

238 B. Dieny, I.L. Prejbeanu, K. Garello, P. Gambardella, P. Freitas, R. Lehndorff, W. Raberg, U. Ebels, S.O. Demokritov, J. Akerman, A. Deac, P. Pirro, C. Adelmann, A. Anane, A.V. Chumak, A. Hirohata, S. Mangin, S.O. Valenzuela, M.C. Onbaşlı, M. d’Aquino, G. Prenat, G. Finocchio, L. Lopez-Diaz, R. Chantrell, O. Chubykalo-Fesenko, and P. Bortolotti, Nature Electronics 3, 446 (2020).

239 D. Apalkov, B. Dieny, and J.M. Slaughter, Proceedings of the IEEE 104, 1796 (2016).

240 S. Lequeux, J. Sampaio, V. Cros, K. Yakushiji, A. Fukushima, R. Matsumoto, H. Kubota, S. Yuasa, and J. Grollier, Scientific Reports 6, (2016).

241 S. Fukami, C. Zhang, S. DuttaGupta, A. Kurenkov, and H. Ohno, Nature Materials 15, 535 (2016).

242 J. Wang, H. Sepehri-Amin, H. Tajiri, T. Nakamura, K. Masuda, Y.K. Takahashi, T. Ina, T. Uruga, I. Suzuki, Y. Miura, and K. Hono, Acta Materialia 166, 413 (2019).

243 K.M. Song, J.-S. Jeong, B. Pan, X. Zhang, J. Xia, S. Cha, T.-E. Park, K. Kim, S. Finizio, J. Raabe, J. Chang, Y. Zhou, W. Zhao, W. Kang, H. Ju, and S. Woo, Nature Electronics 3, 148 (2020).

244 E. Raymenants, A. Vaysset, D. Wan, M. Manfrini, O. Zografos, O. Bultynck, J. Doevenspeck, M. Heyns, I.P. Radu, and T. Devolder, Journal of Applied Physics 124, (2018).

245 M. Prezioso, M.R. Mahmoodi, F.M. Bayat, H. Nili, H. Kim, A. Vincent, and D.B. Strukov, Nature Communications 9, (2018).

246 F.-X. Liang, I.-T. Wang, and T.-H. Hou, Advanced Intelligent Systems 3, 2100007 (2021).

247 J. Torrejon, M. Riou, F.A. Araujo, S. Tsunegi, G. Khalsa, D. Querlioz, P. Bortolotti, V. Cros, K. Yakushiji, A. Fukushima, H. Kubota, S. Yuasa, M.D. Stiles, and J. Grollier, Nature 547, 428 (2017).

248 K. Hayakawa, S. Kanai, T. Funatsu, J. Igarashi, B. Jinnai, W.A. Borders, H. Ohno, and S. Fukami, Physical Review Letters 126, (2021).

249 W.A. Borders, A.Z. Pervaiz, S. Fukami, K.Y. Camsari, H. Ohno, and S. Datta, Nature 573, 390 (2019).

250 M.-H. Wu, M.-C. Hong, C.-C. Chang, P. Sahu, J.-H. Wei, H.-Y. Lee, S.-S. Shcu, and T.-H. Hou, in 2019 Symposium on VLSI Technology (IEEE, 2019).

251 M.-H. Wu, I.-T. Wang, M.-C. Hong, K.-M. Chen, Y.-C. Tseng, J.-H. Wei, and T.-H. Hou, Physical Review Applied 18, (2022).

252 M. Schott, A. Bernand-Mantel, L. Ranno, S. Pizzini, J. Vogel, H. Béa, C. Baraduc, S. Auffret, G. Gaudin, and D. Givord, Nano Letters 17, (2017).

253 C.E. Fillion, J. Fischer, R. Kumar, A. Fassatoui, S. Pizzini, L. Ranno, D. Ourdani, M. Belmeguenai, Y. Roussigné, S.M. Chérif, S. Auffret, I. Joumard, O. Boulle, G. Gaudin, L. Buda-Prejbeanu, C. Baraduc, and H. Béa, Nat Commun 13, 5257 (2022).

254 A. Jaiswal, S. Roy, G. Srinivasan, and K. Roy, IEEE Transactions on Electron Devices 64, 1818 (2017).

255 M. Zahedinejad, H. Fulara, R. Khymyn, A. Houshang, M. Dvornik, S. Fukami, S. Kanai, H. Ohno, and J. Åkerman, Nat Mater 21, 81 (2022).

256 X.-G. Zhang and W.H. Butler, Physical Review B 70, (2004).

257 A. Ross, N. Leroux, A. De Riz, D. Marković, D. Sanz-Hernández, J. Trastoy, P. Bortolotti, D. Querlioz, L. Martins, L. Benetti, M.S. Claro, P. Anacleto, A. Schulman, T. Taris, J.-B. Begueret, S. Saïghi, A.S. Jenkins, R. Ferreira, A.F. Vincent, F.A. Mizrahi, and J. Grollier, Nature Nanotechnology 18, (2023).

258 S. Kanai, K. Hayakawa, H. Ohno, and S. Fukami, Physical Review B 103, (2021).

259 M.-H. Wu, M.-S. Huang, Z. Zhu, F.-X. Liang, M.-C. Hong, J. Deng, J.-H. Wei, S.-S. Sheu, C.-I. Wu, G. Liang, and T.-H. Hou, in 2020 IEEE Symposium on VLSI Technology (IEEE, 2020).

260 J. Choe, in 2023 IEEE International Memory Workshop (IMW) (IEEE, 2023).

261 R. Khymyn, I. Lisenkov, J. Voorheis, O. Sulymenko, O. Prokopenko, V. Tiberkevich, J. Akerman, and A. Slavin, Sci Rep 8, 15727 (2018).

262 F. Trier, P. Noël, J.-V. Kim, J.-P. Attané, L. Vila, and M. Bibes, Nature Reviews Materials 7, (2021).

263 S.S.P. Parkin, M. Hayashi, and L. Thomas, Science 320, (2008).

264 A. Fert, V. Cros, and J. Sampaio, Nature Nanotechnology 8, (2013).

265 M.S. El Hadri, P. Pirro, C.-H. Lambert, S. Petit-Watelot, Y. Quessab, M. Hehn, F. Montaigne, G. Malinowski, and S. Mangin, Physical Review B 94, (2016).

266 A. Fernández-Pacheco, R. Streubel, O. Fruchart, R. Hertel, P. Fischer, and R.P. Cowburn, Nat Commun 8, 15756 (2017).

267 Y. Shen, N.C. Harris, S. Skirlo, M. Prabhu, T. Baehr-Jones, M. Hochberg, X. Sun, S. Zhao, H. Larochelle, D. Englund, and M. Soljačić, Nature Photonics 11, 441 (2017).

268 S. Bandyopadhyay, A. Sludds, S. Krastanov, R. Hamerly, N. Harris, D. Bunandar, M. Streshinsky, M. Hochberg, and D. Englund, ArXiv:2208.01623 [Cs.ET] (2022).

269 J. Feldmann, N. Youngblood, C.D. Wright, H. Bhaskaran, and W.H.P. Pernice, Nature 569, 208 (2019).

270 J. Zhang, S. Dai, Y. Zhao, J. Zhang, and J. Huang, Advanced Intelligent Systems 2, 1900136 (2020).

M. Kumar, S. Abbas, and J. Kim, ACS Applied Materials & Interfaces 10, 34370 (2018).

272 I. Taghavi, M. Moridsadat, A. Tofini, S. Raza, N.A.F. Jaeger, L. Chrostowski, B.J. Shastri, and S. Shekhar, Nanophotonics 11, 3855 (2022).

273 J. Wu, Z.T. Xie, Y. Sha, H.Y. Fu, and Q. Li, Photonics Research 9, 1616 (2021).

274 J. Robertson, P. Kirkland, J.A. Alanis, M. Hejda, J. Bueno, G.D. Caterina, and A. Hurtado, Scientific Reports 12, (2022).

275 S.T. Ilie, J. Faneca, I. Zeimpekis, T.D. Bucio, K. Grabska, D.W. Hewak, H.M.H. Chong, and F.Y. Gardes, Scientific Reports 12, (2022).

276 M. Wang, S. Cai, C. Pan, C. Wang, X. Lian, Y. Zhuo, K. Xu, T. Cao, X. Pan, B. Wang, S.-J. Liang, J.J. Yang, P. Wang, and F. Miao, Nature Electronics 1, 130 (2018).

277 Y. Shi, X. Liang, B. Yuan, V. Chen, H. Li, F. Hui, Z. Yu, F. Yuan, E. Pop, H.-S.P. Wong, and M. Lanza, Nature Electronics 1, 458 (2018).

278 L. Sun, Y. Zhang, G. Han, G. Hwang, J. Jiang, B. Joo, K. Watanabe, T. Taniguchi, Y.-M. Kim, W.J. Yu, B.-S. Kong, R. Zhao, and H. Yang, Nature Communications 10, (2019).

279 F.M. Puglisi, L. Larcher, C. Pan, N. Xiao, Y. Shi, F. Hui, and M. Lanza, in 2016 IEEE International Electron Devices Meeting (IEDM) (IEEE, 2016).

280 X. Yan, J.H. Qian, V.K. Sangwan, and M.C. Hersam, Advanced Materials 34, 2108025 (2022).

281 https://www.fujitsu.com/jp/group/fsm/en/products/reram, Website.

282 S. Chen, M.R. Mahmoodi, Y. Shi, C. Mahata, B. Yuan, X. Liang, C. Wen, F. Hui, D. Akinwande, D.B. Strukov, and M. Lanza, Nature Electronics 3, 638 (2020).

283 K. Lu, X. Li, Q. Sun, X. Pang, J. Chen, T. Minari, X. Liu, and Y. Song, Materials Horizons 8, 447 (2021).

284 S. Pazos, X. Xu, T. Guo, K. Zhu, H.N. Alshareef, and M. Lanza, Nature Reviews Materials 9, (2024).

285 T. Mikolajick, M.H. Park, L. Begon‐Lours, and S. Slesazeck, Advanced Materials 35, (2023).

286 F. Xue, C. Zhang, Y. Ma, Y. Wen, X. He, B. Yu, and X. Zhang, Advanced Materials 34, 2201880 (2022).

287 C. Kaspar, B.J. Ravoo, W.G. van der Wiel, S.V. Wegner, and W.H.P. Pernice, Nature 594, (2021).

288 A. Chen, in 71st Device Research Conference (IEEE, 2013).

289 Y. Shen, W. Zheng, K. Zhu, Y. Xiao, C. Wen, Y. Liu, X. Jing, and M. Lanza, Advanced Materials 33, 2103656 (2021).

290 B. Tang, H. Veluri, Y. Li, Z.G. Yu, M. Waqar, J.F. Leong, M. Sivan, E. Zamburg, Y.-W. Zhang, J. Wang, and A.V.-Y. Thean, Nature Communications 13, (2022).

291 X. Xu, T. Guo, H. Kim, M.K. Hota, R.S. Alsaadi, M. Lanza, X. Zhang, and H.N. Alshareef, Advanced Materials 34, (2022).

292 T.-A. Chen, C.-P. Chuu, C.-C. Tseng, C.-K. Wen, H.-S.P. Wong, S. Pan, R. Li, T.-A. Chao, W.-C. Chueh, Y. Zhang, Q. Fu, B.I. Yakobson, W.-H. Chang, and L.-J. Li, Nature 579, 219 (2020).

293 K.Y. Ma, L. Zhang, S. Jin, Y. Wang, S.I. Yoon, H. Hwang, J. Oh, D.S. Jeong, M. Wang, S. Chatterjee, G. Kim, A.-R. Jang, J. Yang, S. Ryu, H.Y. Jeong, R.S. Ruoff, M. Chhowalla, F. Ding, and H.S. Shin, Nature 606, 88 (2022).

294 G. Migliato Marega, H.G. Ji, Z. Wang, G. Pasquale, M. Tripathi, A. Radenovic, and A. Kis, Nature Electronics 6, (2023).

295 M. Lanza, R. Waser, D. Ielmini, J.J. Yang, L. Goux, J. Suñe, A.J. Kenyon, A. Mehonic, S. Spiga, V. Rana, S. Wiefels, S. Menzel, I. Valov, M.A. Villena, E. Miranda, X. Jing, F. Campabadal, M.B. Gonzalez, F. Aguirre, F. Palumbo, K. Zhu, J.B. Roldan, F.M. Puglisi, L. Larcher, T.-H. Hou, T. Prodromakis, Y. Yang, P. Huang, T. Wan, Y. Chai, K.L. Pey, N. Raghavan, S. Dueñas, T. Wang, Q. Xia, and S. Pazos, ACS Nano 15, 17214 (2021).

296 M. Lanza, A. Sebastian, W.D. Lu, G.M. Le, M.F. Chang, D. Akinwande, F.M. Puglisi, H.N. Alshareef, M. Liu, and J.B. Roldan, Science 376, 9979 (2022).

297 D. Edelstein, M. Rizzolo, D. Sil, A. Dutta, J. DeBrosse, M. Wordeman, A. Arceo, I.C. Chu, J. Demarest, E.R.J. Edwards, E.R. Evarts, J. Fullam, A. Gasasira, G. Hu, M. Iwatake, R. Johnson, V. Katragadda, T. Levin, J. Li, Y. Liu, C. Long, T. Maffitt, S. McDermott, S. Mehta, V. Mehta, D. Metzler, J. Morillo, Y. Nakamura, S. Nguyen, P. Nieves, V. Pai, R. Patlolla, R. Pujari, R. Southwick, T. Standaert, O. van der Straten, H. Wu, C.-C. Yang, D. Houssameddine, J.M. Slaughter, and D.C. Worledge, in 2020 IEEE International Electron Devices Meeting (IEDM) (IEEE, 2020).

298 F. Xiong, E. Yalon, A. Behnam, C.M. Neumann, K.L. Grosse, S. Deshmukh, and E. Pop, in 2016 IEEE International Electron Devices Meeting (IEDM) (IEEE, 2016).

299 M.L. Gallo and A. Sebastian, Journal of Physics D: Applied Physics 53, 213002 (2020).

300 R. Dittmann, S. Menzel, and R. Waser, Advances in Physics 70, 155 (2021).

301 V.V. Zhirnov, R. Meade, R.K. Cavin, and G. Sandhu, Nanotechnology 22, 254027 (2011).

302 E. Chicca and G. Indiveri, Applied Physics Letters 116, 120501 (2020).

303 C. Safranski, G. Hu, J.Z. Sun, P. Hashemi, S.L. Brown, L. Buzi, C.P. D'Emic, E.R.J. Edwards, E. Galligan, M.G. Gottwald, O. Gunawan, S. Karimeddiny, H. Jung, J. Kim, K. Latzko, P.L. Trouilloud, S. Zare, and D.C. Worledge, in 2022 IEEE Symposium on VLSI Technology and Circuits (VLSI Technology and Circuits) (IEEE, 2022).

304 C. Persch, M.J. Müller, A. Yadav, J. Pries, N. Honné, P. Kerres, S. Wei, H. Tanaka, P. Fantini, E. Varesi, F. Pellizzer, and M. Wuttig, Nature Communications 12, (2021).

305 M. von Witzleben, S. Wiefels, A. Kindsmüller, P. Stasner, F. Berg, F. Cüppers, S. Hoffmann-Eifert, R. Waser, S. Menzel, and U. Böttger, ACS Applied Electronic Materials 3, 5563 (2021).

306 S. Wiefels, M.V. Witzleben, M. Huttemann, U. Bottger, R. Waser, and S. Menzel, IEEE Transactions on Electron Devices 68, 1024 (2021).

N. Kopperberg, S. Wiefels, S. Liberda, R. Waser, and S. Menzel, ACS Applied Materials & Interfaces 13, 58066 (2021).

308 C. Bengel, J. Mohr, S. Wiefels, A. Singh, A. Gebregiorgis, R. Bishnoi, S. Hamdioui, R. Waser, D. Wouters, and S. Menzel, Neuromorphic Computing and Engineering 2, 034001 (2022).

309 K. Schnieders, C. Funck, F. Cüppers, S. Aussen, T. Kempen, A. Sarantopoulos, R. Dittmann, S. Menzel, V. Rana, S. Hoffmann-Eifert, and S. Wiefels, APL Materials 10, 101114 (2022).

310 S. Slesazeck, V. Havel, E. Breyer, H. Mulaosmanovic, M. Hoffmann, B. Max, S. Duenkel, and T. Mikolajick, in 2019 IEEE 11th International Memory Workshop (IMW) (IEEE, 2019).

311 J.H. Yoon, S.J. Song, I.H. Yoo, J.Y. Seok, K.J. Yoon, D.E. Kwon, T.H. Park, and C.S. Hwang, Advanced Functional Materials 24, (2014).

312 M. Lanza, H.-S.P. Wong, E. Pop, D. Ielmini, D. Strukov, B.C. Regan, L. Larcher, M.A. Villena, J.J. Yang, L. Goux, A. Belmonte, Y. Yang, F.M. Puglisi, J. Kang, B. Magyari-Köpe, E. Yalon, A. Kenyon, M. Buckwell, A. Mehonic, A. Shluger, H. Li, T.-H. Hou, B. Hudec, D. Akinwande, R. Ge, S. Ambrogio, J.B. Roldan, E. Miranda, J. Suñe, K.L. Pey, X. Wu, N. Raghavan, E. Wu, W.D. Lu, G. Navarro, W. Zhang, H. Wu, R. Li, A. Holleitner, U. Wurstbauer, M.C. Lemme, M. Liu, S. Long, Q. Liu, H. Lv, A. Padovani, P. Pavan, I. Valov, X. Jing, T. Han, K. Zhu, S. Chen, F. Hui, and Y. Shi, Advanced Electronic Materials 5, 1800143 (2018).

313 G. Dearnaley, A.M. Stoneham, and D.V. Morgan, Reports on Progress in Physics 33, 1129 (1970).

314 Y. Yang, P. Gao, S. Gaba, T. Chang, X. Pan, and W. Lu, Nature Communications 3, (2012).

315 Y. Yang, P. Gao, L. Li, X. Pan, S. Tappertzhofen, S.H. Choi, R. Waser, I. Valov, and W.D. Lu, Nature Communications 5, (2014).

316 R. Waser, R. Dittmann, G. Staikov, and K. Szot, Advanced Materials 21, 2632 (2009).

317 S. Meister, S.B. Kim, J.J. Cha, H.-S.P. Wong, and Y. Cui, ACS Nano 5, (2011).

318 S. Chen, L. Jiang, M. Buckwell, X. Jing, Y. Ji, E. Grustan-Gutierrez, F. Hui, Y. Shi, M. Rommel, A. Paskaleva, G. Benstetter, W.H. Ng, A. Mehonic, A.J. Kenyon, and M. Lanza, Advanced Functional Materials 28, 1802266 (2018).

319 M. Buckwell, L. Montesi, S. Hudziak, A. Mehonic, and A.J. Kenyon, Nanoscale 7, 18030 (2015).

320 A. Mehonic, M. Buckwell, L. Montesi, M.S. Munde, D. Gao, S. Hudziak, R.J. Chater, S. Fearn, D. McPhail, M. Bosman, A.L. Shluger, and A.J. Kenyon, Advanced Materials 28, 7486 (2016).

321 K. Szot, W. Speier, G. Bihlmayer, and R. Waser, Nature Materials 5, 312 (2006).

322 H.R.J. Cox, M. Buckwell, W.H. Ng, D.J. Mannion, A. Mehonic, P.R. Shearing, S. Fearn, and A.J. Kenyon, APL Materials 9, 111109 (2021).

323 I. Valov and T. Tsuruoka, Journal of Physics D: Applied Physics 51, 413001 (2018).

324 A. Mehonic, A. Vrajitoarea, S. Cueff, S. Hudziak, H. Howe, C. Labbé, R. Rizk, M. Pepper, and A.J. Kenyon, Scientific Reports 3, (2013).

325 S. Stathopoulos, L. Michalas, A. Khiat, A. Serb, and T. Prodromakis, Scientific Reports 9, (2019).

326 R. Yuan, Q. Duan, P.J. Tiw, G. Li, Z. Xiao, Z. Jing, K. Yang, C. Liu, C. Ge, R. Huang, and Y. Yang, Nature Communications 13, (2022).

327 J. del Valle, P. Salev, F. Tesler, N.M. Vargas, Y. Kalcheim, P. Wang, J. Trastoy, M.-H. Lee, G. Kassabian, J.G. Ramírez, M.J. Rozenberg, and I.K. Schuller, Nature 569, 388 (2019).

328 S.G. Sarwat, B. Kersting, T. Moraitis, V.P. Jonnalagadda, and A. Sebastian, Nature Nanotechnology 17, 507 (2022).

329 G.W. Burr, M.J. BrightSky, A. Sebastian, H.-Y. Cheng, J.-Y. Wu, S. Kim, N.E. Sosa, N. Papandreou, H.-L. Lung, H. Pozidis, E. Eleftheriou, and C.H. Lam, IEEE Journal on Emerging and Selected Topics in Circuits and Systems 6, (2016).

330 M. Salinga, B. Kersting, I. Ronneberger, V.P. Jonnalagadda, X.T. Vu, G.M. Le, I. Giannopoulos, O. Cojocaru-Mirédin, R. Mazzarello, and A. Sebastian, Nat Mater 17, 681 (2018).

331 F. Rao, K. Ding, Y. Zhou, Y. Zheng, M. Xia, S. Lv, Z. Song, S. Feng, I. Ronneberger, R. Mazzarello, W. Zhang, and E. Ma, Science 358, 1423 (2017).

332 S. Bhatti, R. Sbiaa, A. Hirohata, H. Ohno, S. Fukami, and S.N. Piramanayagam, Materials Today 20, (2017).

333 L.W. Martin and A.M. Rappe, Nature Reviews Materials 2, (2016).

334 J. Hwang, Y. Goh, and S. Jeon, Small 20, (2023).

335 K.-H. Kim, I. Karpov, R.H. Olsson, and D. Jariwala, Nature Nanotechnology 18, (2023).

336 A. Chanthbouala, V. Garcia, R.O. Cherifi, K. Bouzehouane, S. Fusil, X. Moya, S. Xavier, H. Yamada, C. Deranlot, N.D. Mathur, M. Bibes, A. Barthélémy, and J. Grollier, Nature Materials 11, 860 (2012).

337 M. Onen, N. Emond, B. Wang, D. Zhang, F.M. Ross, J. Li, B. Yildiz, and J.A. del Alamo, Science 377, 539 (2022).

338 M.L. Gallo, T. Tuma, F. Zipoli, A. Sebastian, and E. Eleftheriou, in 2016 46th European Solid-State Device Research Conference (ESSDERC) (IEEE, 2016).

339 A. Fukushima, T. Seki, K. Yakushiji, H. Kubota, H. Imamura, S. Yuasa, and K. Ando, Applied Physics Express 7, 083001 (2014).

340 D. Vodenicarevic, N. Locatelli, A. Mizrahi, J.S. Friedman, A.F. Vincent, M. Romera, A. Fukushima, K. Yakushiji, H. Kubota, S. Yuasa, S. Tiwari, J. Grollier, and D. Querlioz, Physical Review Applied 8, (2017).

341 W. Qian, X. Li, M.D. Riedel, K. Bazargan, and D.J. Lilja, IEEE Transactions on Computers 60, (2011).

342 F. Neugebauer, I. Polian, and J.P. Hayes, in 2017 Euromicro Conference on Digital System Design (DSD) (IEEE, 2017).

343 K.-E. Harabi, T. Hirtzlin, C. Turck, E. Vianello, R. Laurent, J. Droulez, P. Bessière, J.-M. Portal, M. Bocquet, and D. Querlioz, Nature Electronics (2022).

344 A. Sengupta, G. Srinivasan, D. Roy, and K. Roy, in 2018 IEEE International Electron Devices Meeting (IEDM) (IEEE, 2018).

345 K.Y. Camsari, R. Faria, B.M. Sutton, and S. Datta, Physical Review X 7, (2017).

346 A. Mizrahi, T. Hirtzlin, A. Fukushima, H. Kubota, S. Yuasa, J. Grollier, and D. Querlioz, Nature Communications 9, (2018).

347 A. Sengupta, P. Panda, P. Wijesinghe, Y. Kim, and K. Roy, Scientific Reports 6, (2016).

348 N. Mohseni, P.L. McMahon, and T. Byrnes, Nature Reviews Physics 4, (2022).

349 A. Grimaldi, K. Selcuk, N.A. Aadit, K. Kobayashi, Q. Cao, S. Chowdhury, G. Finocchio, S. Kanai, H. Ohno, S. Fukami, and K.Y. Camsari, in 2022 International Electron Devices Meeting (IEDM) (IEEE, 2022).

350 N.A. Aadit, A. Grimaldi, M. Carpentieri, L. Theogarajan, J.M. Martinis, G. Finocchio, and K.Y. Camsari, Nature Electronics 5, (2022).

351 C.-Y. Huang, W.C. Shen, Y.-H. Tseng, Y.-C. King, and C.-J. Lin, IEEE Electron Device Letters 33, 1108 (2012).

352 H. Nili, G.C. Adam, B. Hoskins, M. Prezioso, J. Kim, M.R. Mahmoodi, F.M. Bayat, O. Kavehei, and D.B. Strukov, Nature Electronics 1, 197 (2018).

353 A.F. Vincent, J. Larroque, N. Locatelli, N.B. Romdhane, O. Bichler, C. Gamrat, W.S. Zhao, J.-O. Klein, S. Galdin-Retailleau, and D. Querlioz, IEEE Transactions on Biomedical Circuits and Systems 9, 166 (2015).

354 J.Z. Sun, Physical Review B 62, 570 (2000).

355 J. Kaiser, W.A. Borders, K.Y. Camsari, S. Fukami, H. Ohno, and S. Datta, Physical Review Applied 17, (2022).

356 A. Sebastian, R. Pendurthi, A. Kozhakhmetov, N. Trainor, J.A. Robinson, J.M. Redwing, and S. Das, Nature Communications 13, (2022).

357 S. Liu, T.P. Xiao, J. Kwon, B.J. Debusschere, S. Agarwal, J.A.C. Incorvia, and C.H. Bennett, Frontiers in Nanotechnology 4, (2022).

358 D. Bonnet, T. Hirtzlin, A. Majumdar, T. Dalgaty, E. Esmanhotto, V. Meli, N. Castellani, S. Martin, J.-F. Nodin, G. Bourgeois, J.-M. Portal, D. Querlioz, and E. Vianello, Nature Communications 14, 7530 (2023).

359 K. Roy, A. Jaiswal, and P. Panda, Nature 575, 607 (2019).

360 B. Wickramasinghe, S.S. Chowdhury, A.K. Kosta, W. Ponghiran, and K. Roy, IEEE Transactions on Cognitive and Developmental Systems 1 (2024).

361 A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A.N. Gomez, Ł. Kaiser, and I. Polosukhin, Advances in Neural Information Processing Systems 30, (2017).

362 T. Mikolov, M. Karafiát, L. Burget, J. Černocký, and S. Khudanpur, in Interspeech 2010 (ISCA, 2010).

363 A. Graves, A.-rahman Mohamed, and G. Hinton, in 2013 IEEE International Conference on Acoustics Speech and Signal Processing (IEEE, 2013).

364 S. Hochreiter and J. Schmidhuber, Neural Comput 9, 1735 (1997).

365in Proceedings of SSST-8, Eighth Workshop on Syntax, Semantics and Structure in Statistical Translation (https://arxiv.org/abs/1409.1259, 2014), pp. 103–111.

366in NIPS 2014 Workshop on Deep Learning (https://arxiv.org/abs/1412.3555, 2014).

367IEEE Signal Processing Magazine 36(6), 51 (2019).

368 S.S. Chowdhury, C. Lee, and K. Roy, Neurocomputing 464, 83 (2021).

369 W. Ponghiran and K. Roy, in Proceedings of the AAAI Conference on Artificial Intelligence (2022), pp. 8001–8008.

370in In International Conference on Learning Representations (https://arxiv.org/abs/2010.11929v2, 2020).

371 P.U. Diehl, D. Neil, J. Binas, M. Cook, S.-C. Liu, and M. Pfeiffer, in 2015 International Joint Conference on Neural Networks (IJCNN) (IEEE, 2015).

372 C. Lee, S.S. Sarwar, P. Panda, G. Srinivasan, and K. Roy, Front Neurosci 14, 119 (2020).

373 J. Kim, H. Kim, S. Huh, J. Lee, and K. Choi, Neurocomputing 311, 373 (2018).

374in Proceedings of the 56th Annual Design Automation Conference (https://doi.org/10.1145/3316781.3317822, 2019), pp. 1–6.

375 I. Garg, S.S. Chowdhury, and K. Roy, in 2021 IEEE/CVF International Conference on Computer Vision (ICCV) (IEEE, 2021).

376in 57th ACM/IEEE Design Automation Conference (DAC) (https://dl.acm.org/doi/10.5555/3437539.3437564, 2020), pp. 1–6.

B. Han and K. Roy, in Computer Vision – ECCV 2020 (Springer International Publishing, 2020), pp. 388–404.

378 N. Rathi and K. Roy, IEEE Trans Neural Netw Learn Syst PP, (2021).

379 B. Rueckauer, I.A. Lungu, Y. Hu, M. Pfeiffer, and S.C. Liu, Front Neurosci 11, 682 (2017).

380 C. Brandli, R. Berner, M. Yang, S.-C. Liu, and T. Delbruck, IEEE Journal of Solid-State Circuits 49, 2333 (2014).

381 A. Zhu, L. Yuan, K. Chaney, and K. Daniilidis, in Robotics: Science and Systems XIV (Robotics: Science and Systems Foundation, 2018).

382 A.Z. Zhu, L. Yuan, K. Chaney, and K. Daniilidis, in 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW) (IEEE, 2019).

383 A.K. Kosta and K. Roy, in IEEE International Conference on Robotics and Automation (ICRA) (https://arxiv.org/abs/2209.11741, 2023), pp. 6021–6027.

384 A. Sengupta, Y. Ye, R. Wang, C. Liu, and K. Roy, Frontiers in Neuroscience 13, (2019).

385 J. Kaiser, H. Mostafa, and E. Neftci, Front Neurosci 14, 424 (2020).

386 D. Huh and T.J. Sejnowski, in Advances in Neural Information Processing Systems (https://arxiv.org/abs/1706.04698, 2018).

387in International Conference on Learning Representations (https://arxiv.org/abs/2005.01807, 2019).

388in European Conference on Computer Vision (Pp. 709-726). Cham: Springer Nature Switzerland (https://arxiv.org/abs/2110.05929, 2022).

389(2022).

390 K. Suetake, S.I. Ikegawa, R. Saiin, and Y. Sawada, Neural Netw 159, 208 (2023).

391 P. O’Connor, E. Gavves, and M. Welling, in The 22nd International Conference on Artificial Intelligence and Statistics (PMLR, 2019), pp. 1516–1523.

392 M. Xiao, Q. Meng, Z. Zhang, Y. Wang, and Z. Lin, Advances in Neural Information Processing Systems 34, 14516 (2021).

393 N. Rathi, I. Chakraborty, A. Kosta, A. Sengupta, A. Ankit, P. Panda, and K. Roy, ACM Computing Surveys 55, 1 (2023).

394 B.V. Benjamin, P. Gao, E. McQuinn, S. Choudhary, A.R. Chandrasekaran, J.-M. Bussat, R. Alvarez-Icaza, J.V. Arthur, P.A. Merolla, and K. Boahen, Proceedings of the IEEE 102, 699 (2014).

395 P.A. Merolla, J.V. Arthur, R. Alvarez-Icaza, A.S. Cassidy, J. Sawada, F. Akopyan, B.L. Jackson, N. Imam, C. Guo, Y. Nakamura, B. Brezzo, I. Vo, S.K. Esser, R. Appuswamy, B. Taba, A. Amir, M.D. Flickner, W.P. Risk, R. Manohar, and D.S. Modha, Science 345, 668 (2014).

396 D.G. Elliott, M. Stumm, W.M. Snelgrove, C. Cojocaru, and R. Mckenzie, IEEE Design & Test of Computers 16, 32 (1999).

397 A. Agrawal, M. Ali, M. Koo, N. Rathi, A. Jaiswal, and K. Roy, IEEE Solid-State Circuits Letters 4, 137 (2021).

398 I. Chakraborty, A. Jaiswal, A.K. Saha, S.K. Gupta, and K. Roy, Applied Physics Reviews 7, 021308 (2020).

399 R. Waser, R. Dittmann, G. Staikov, and K. Szot, Adv Mater 21, 2632 (2009).

400 G.W. Burr, M.J. BrightSky, A. Sebastian, H.-Y. Cheng, J.-Y. Wu, S. Kim, N.E. Sosa, N. Papandreou, H.-L. Lung, H. Pozidis, E. Eleftheriou, and C.H. Lam, IEEE Journal on Emerging and Selected Topics in Circuits and Systems 6, 146 (2016).

401 M. Hosomi, H. Yamagishi, T. Yamamoto, K. Bessho, Y. Higo, K. Yamane, H. Yamada, M. Shoji, H. Hachino, C. Fukumoto, H. Nagao, and H. Kano, in IEEE InternationalElectron Devices Meeting 2005. IEDM Technical Digest. (IEEE, n.d.).

402 A. Sengupta and K. Roy, Applied Physics Reviews 4, 041105 (2017).

403 T. Sharma, C. Wang, A. Agrawal, and K. Roy, in 2021 IEEE/ACM International Symposium on Low Power Electronics and Design (ISLPED) (IEEE, 2021).

404 F. Merrikh-Bayat, X. Guo, M. Klachko, M. Prezioso, K.K. Likharev, and D.B. Strukov, IEEE Trans Neural Netw Learn Syst 29, 4782 (2018).

405 P. Panda and K. Roy, in International Joint Conference on Neural Networks (IJCNN) (https://arxiv.org/abs/1602.01510, 2016), pp. 299–306.

406 Z. Sun and D. Ielmini, IEEE Transactions on Circuits and Systems II: Express Briefs 69, 3024 (2022).

407 Z. Sun, S. Kvatinsky, X. Si, A. Mehonic, Y. Cai, and R. Huang, Nature Electronics 6, (2023).

408 M. Le Gallo, A. Sebastian, R. Mathis, M. Manica, H. Giefers, T. Tuma, C. Bekas, A. Curioni, and E. Eleftheriou, Nature Electronics 1, 246 (2018).

409 M. Le Gallo, A. Sebastian, G. Cherubini, H. Giefers, and E. Eleftheriou, IEEE Transactions on Electron Devices 65, (2018).

410 E.J. Fuller, S.T. Keene, A. Melianas, Z. Wang, S. Agarwal, Y. Li, Y. Tuchman, C.D. James, M.J. Marinella, J.J. Yang, A. Salleo, and A.A. Talin, Science 364, 570 (2019).

411 T.P. Xiao, B. Feinberg, C.H. Bennett, V. Agrawal, P. Saxena, V. Prabhakar, K. Ramkumar, H. Medu, V. Raghavan, R. Chettuvetty, S. Agarwal, and M.J. Marinella, IEEE Transactions on Circuits and Systems I: Regular Papers 69, (2022).

412.

413 Y. Luo, S. Wang, P. Zuo, Z. Sun, and R. Huang, IEEE Transactions on Circuits and Systems I: Regular Papers 1 (2022).

414 Z. Sun, G. Pedretti, E. Ambrosi, A. Bricalli, W. Wang, and D. Ielmini, Proceedings of the National Academy of Sciences 116, 4123 (2019).

415 B. Feinberg, R. Wong, T.P. Xiao, C.H. Bennett, J.N. Rohan, E.G. Boman, M.J. Marinella, S. Agarwal, and E. Ipek, in 2021 IEEE International Symposium on High-Performance Computer Architecture (HPCA) (IEEE, 2021).

416 P. Mannocci, E. Melacarne, A. Pezzoli, G. Pedretti, C. Villa, F. Sancandi, U. Spagnolini, and D. Ielmini, in 2023 International Electron Devices Meeting (IEDM) (IEEE, 2023).

417 Q. Zeng, J. Liu, M. Jiang, J. Lan, Y. Gong, Z. Wang, Y. Li, C. Li, J. Ignowski, and K. Huang, IEEE Internet of Things Journal 11, (2024).

418 S. Wang, Y. Luo, P. Zuo, L. Pan, Y. Li, and Z. Sun, Science Advances 9, (2023).

419 Z. Chen, Y. Zhou, H. Xu, Y. Fu, Y. Li, Y. He, X.-S. Miao, P. Mannocci, and D. Ielmini, in 2023 International Electron Devices Meeting (IEDM) (IEEE, 2023).

420 J. Woo and S. Yu, IEEE Nanotechnology Magazine 12, (2018).

421 S. Achour, R. Sarpeshkar, and M.C. Rinard, ACM SIGPLAN Notices 51, 177 (2016).

422 J. Zhao, S. Huang, O. Yousuf, Y. Gao, B.D. Hoskins, and G.C. Adam, Frontiers in Neuroscience 15, (2021).

423 W. Yi, F. Perner, M.S. Qureshi, H. Abdalla, M.D. Pickett, J.J. Yang, M.-X.M. Zhang, G. Medeiros-Ribeiro, and R.S. Williams, Applied Physics A 102, (2011).

424.

425 B. Feinberg, U.K.R. Vengalam, N. Whitehair, S. Wang, and E. Ipek, in 2018 ACM/IEEE 45th Annual International Symposium on Computer Architecture (ISCA) (IEEE, 2018).

426 P. Zuo, Z. Sun, and R. Huang, IEEE Transactions on Circuits and Systems II: Express Briefs 70, (2023).

427 P. Mannocci, E. Giannone, and D. Ielmini, in 2023 IEEE International Conference on Metrology for EXtended Reality, Artificial Intelligence and Neural Engineering (MetroXRAINE) (IEEE, 2023).

428 P. Mannocci and D. Ielmini, IEEE Journal on Exploratory Solid-State Computational Devices and Circuits 9, (2023).

429 S. Shapero, A.S. Charles, C.J. Rozell, and P. Hasler, IEEE Journal on Emerging and Selected Topics in Circuits and Systems 2, 530 (2012).

430 S. Rehman, M.F. Khan, H.-D. Kim, and S. Kim, Nature Communications 13, (2022).

431 C.E. Graves, C. Li, G. Pedretti, and J.P. Strachan, in Memristor Computing Systems (Springer International Publishing, 2022), pp. 105–139.

432 C. Li, C.E. Graves, X. Sheng, D. Miller, M. Foltin, G. Pedretti, and J.P. Strachan, Nature Communications 11, (2020).

433 X.S. Hu, M. Niemier, A. Kazemi, A.F. Laguna, K. Ni, R. Rajaei, M.M. Sharifi, and X. Yin, in 2021 IEEE International Electron Devices Meeting (IEDM) (IEEE, 2021).

434 L. Grinsztajn, E. Oyallon, and G. Varoquaux, NeurIPS 2022 Datasets and Benchmarks Track (2022).

435 Z. Xie, W. Dong, J. Liu, H. Liu, and D. Li, in Proceedings of the Sixteenth European Conference on Computer Systems (ACM, 2021).

436 G. Pedretti, C.E. Graves, S. Serebryakov, R. Mao, X. Sheng, M. Foltin, C. Li, and J.P. Strachan, Nature Communications 12, (2021).

437 A. Shafiee, A. Nag, N. Muralimanohar, R. Balasubramonian, J.P. Strachan, M. Hu, R.S. Williams, and V. Srikumar, in 2016 ACM/IEEE 43rd Annual International Symposium on Computer Architecture (ISCA) (IEEE, 2016).

438 G. Pedretti, J. Moon, P. Bruel, S. Serebryakov, R.M. Roth, L. Buonanno, A. Gajjar, T. Ziegler, C. Xu, M. Foltin, P. Faraboschi, J. Ignowski, and C.E. Graves, ArXiv (2023).

439 K. Pagiamtzis and A. Sheikholeslami, IEEE Journal of Solid-State Circuits 41, 712 (2006).

440 G. Pedretti, C.E. Graves, T.V. Vaerenbergh, S. Serebryakov, M. Foltin, X. Sheng, R. Mao, C. Li, and J.P. Strachan, Advanced Electronic Materials 8, 2101198 (2022).

441 H. Farzaneh, J.P.C. de Lima, M. Li, A.A. Khan, X.S. Hu, and J. Castrillon, ArXiv (2023).

442 A. Bremler-Barr, D. Hay, D. Hendler, and R.M. Roth, IEEE/ACM Transactions on Networking 18, 1665 (2010).

443 A. Lucas, Frontiers in Physics 2, (2014).

444 J.J. Hopfield and D.W. Tank, Biological Cybernetics 52, 141 (1985).

445 E. Farhi, J. Goldstone, and S. Gutmann, (2014).

446 M. Ercsey-Ravasz and Z. Toroczkai, Nature Physics 7, 966 (2011).

447 B. Molnár, F. Molnár, M. Varga, Z. Toroczkai, and M. Ercsey-Ravasz, Nature Communications 9, (2018).

448 T.J. Sejnowski, in AIP Conference Proceedings (AIP, 1986).

449 R. Hamerly, T. Inagaki, P.L. McMahon, D. Venturelli, A. Marandi, T. Onodera, E. Ng, C. Langrock, K. Inaba, T. Honjo, and others, Feedback 1, 2 (2018).

450 T.-J. Yang and V. Sze, in 2019 IEEE International Electron Devices Meeting (IEDM) (IEEE, 2019), pp. 22–1.

451 J. Deguchi, D. Miyashita, A. Maki, S. Sasaki, K. Nakata, and F. Tachibana, in 2019 IEEE International Electron Devices Meeting (IEDM) (IEEE, 2019), pp. 22–4.

452 H. Jia, M. Ozatay, Y. Tang, H. Valavi, R. Pathak, J. Lee, and N. Verma, IEEE Journal of Solid-State Circuits 57, 198 (2022).

453 L. Fick, S. Skrzyniarz, M. Parikh, M.B. Henry, and D. Fick, in 2022 IEEE International Solid- State Circuits Conference (ISSCC) (IEEE, 2022).

454 P. Narayanan, S. Ambrogio, A. Okazaki, K. Hosokawa, H. Tsai, A. Nomura, T. Yasuda, C. Mackin, S.C. Lewis, A. Friz, M. Ishii, Y. Kohda, H. Mori, K. Spoon, R. Khaddam-Aljameh, N. Saulnier, M. Bergendahl, J. Demarest, K.W. Brew, V. Chan, S. Choi, I. Ok, I. Ahsan, F.L. Lie, W. Haensch, V. Narayanan, and G.W. Burr, IEEE Transactions on Electron Devices 68, 6629 (2021).

455 M.L. Gallo, R. Khaddam-Aljameh, M. Stanisavljevic, A. Vasilopoulos, B. Kersting, M. Dazzi, G. Karunaratne, M. Brändli, A. Singh, S.M. Müller, J. Büchel, X. Timoneda, V. Joshi, M.J. Rasch, U. Egger, A. Garofalo, A. Petropoulos, T. Antonakopoulos, K. Brew, S. Choi, I. Ok, T. Philip, V. Chan, C. Silvestre, I. Ahsan, N. Saulnier, V. Narayanan, P.A. Francese, E. Eleftheriou, and A. Sebastian, Nature Electronics (2023).

456 S. Kim, M. Ishii, S. Lewis, T. Perri, M. BrightSky, W. Kim, R. Jordan, G.W. Burr, N. Sosa, A. Ray, J.-P. Han, C. Miller, K. Hosokawa, and C. Lam, in 2015 IEEE International Electron Devices Meeting (IEDM) (IEEE, 2015).

457 S. Kim, S. Kim, S. Um, S. Kim, K. Kim, and H.-J. Yoo, in 2022 IEEE Symposium on VLSI Technology and Circuits (VLSI Technology and Circuits) (IEEE, 2022).

458 M. Anderson, B. Chen, S. Chen, S. Deng, J. Fix, M. Gschwind, A. Kalaiah, C. Kim, J. Lee, J. Liang, and et al, ArXiv Preprint ArXiv:2107.04140 (2021).

459 R. Khaddam-Aljameh, M. Stanisavljevic, J.F. Mas, G. Karunaratne, M. Brandli, F. Liu, A. Singh, S.M. Muller, U. Egger, A. Petropoulos, T. Antonakopoulos, K. Brew, S. Choi, I. Ok, F.L. Lie, N. Saulnier, V. Chan, I. Ahsan, V. Narayanan, S.R. Nandakumar, M.L. Gallo, P.A. Francese, A. Sebastian, and E. Eleftheriou, IEEE Journal of Solid-State Circuits 57, 1027 (2022).

460 S. Jain, H. Tsai, C.-T. Chen, R. Muralidhar, I. Boybat, M.M. Frank, S. Wozniak, M. Stanisavljevic, P. Adusumilli, P. Narayanan, K. Hosokawa, M. Ishii, A. Kumar, V. Narayanan, and G.W. Burr, IEEE Transactions on Very Large Scale Integration (VLSI) Systems 31, 114 (2023).

461 A.I. Khan, A. Daus, R. Islam, K.M. Neilson, H.R. Lee, H.P. Wong, and E. Pop, Science 373, 1243 (2021).

462 B. Prasad, S. Parkin, T. Prodromakis, C.-B. Eom, J. Sort, and J.L. MacManus-Driscoll, APL Materials 10, (2022).

463 G.W. Burr, A. Sebastian, E. Vianello, R. Waser, and S. Parkin, APL Materials 8, (2020).

464 T. Venkatesan and S. Williams, Applied Physics Reviews 9, (2022).

465 T. Chen, J. van Gelder, B. van de Ven, S.V. Amitonov, B. de Wilde, H.-C. Ruiz Euler, H. Broersma, P.A. Bobbert, F.A. Zwanenburg, and W.G. van der Wiel, Nature 577, (2020).

466 R.A. John, Y. Demirağ, Y. Shynkarenko, Y. Berezovska, N. Ohannessian, M. Payvand, P. Zeng, M.I. Bodnarchuk, F. Krumeich, G. Kara, I. Shorubalko, M.V. Nair, G.A. Cooke, T. Lippert, G. Indiveri, and M.V. Kovalenko, Nature Communications 13, (2022).

467 N. Yantara, S.E. Ng, D. Sharma, B. Zhou, P.S.V. Sun, H.M. Chua, N.F. Jamaludin, A. Basu, and N. Mathews, Advanced Materials 36, (2023).

468 E.J. Fuller, S.T. Keene, A. Melianas, Z. Wang, S. Agarwal, Y. Li, Y. Tuchman, C.D. James, M.J. Marinella, J.J. Yang, A. Salleo, and A.A. Talin, Science 364, (2019).

469 Y. van de Burgt, E. Lubberman, E.J. Fuller, S.T. Keene, G.C. Faria, S. Agarwal, M.J. Marinella, A. Alec Talin, and A. Salleo, Nature Materials 16, (2017).

470 S. Goswami, R. Pramanick, A. Patra, S.P. Rath, M. Foltin, A. Ariando, D. Thompson, T. Venkatesan, S. Goswami, and R.S. Williams, Nature 597, (2021).

471 S. Goswami, A.J. Matula, S.P. Rath, S. Hedström, S. Saha, M. Annamalai, D. Sengupta, A. Patra, S. Ghosh, H. Jani, S. Sarkar, M.R. Motapothula, C.A. Nijhuis, J. Martin, S. Goswami, V.S. Batista, and T. Venkatesan, Nature Materials 16, (2017).

472 M. Lanza, G. Molas, and I. Naveh, Nature Electronics 6, (2023).