Matter
PerspectiveAccelerate and actualize: Can 2D materials bridge the gap between neuromorphic hardware and the human brain?
Progress and potential
2D materials present an exciting opportunity for devices and systems beyond the von Neumann computing architecture paradigm due to their diversity of electronic structure, physical properties, and atomically thin van der Waals structures that enable ease of integration with conventional electronic materials and Si-based hardware. All major classes of non-volatile memory (NVM) devices have been demonstrated using 2D materials, including their operation as synaptic devices for applications in neuromorphic computing hardware. Their atomically thin structure and superior physical properties, i.e., mechanical strength and electrical and thermal conductivity, as well as their gate-tunable electronic properties provide performance advantages and novel functionality in NVM devices and systems. However, device performance and variability compared with incumbent materials and technology remain major concerns for real applications. Ultimately, the progress of 2D materials as a novel class of electronic materials, and specifically their application in the area of neuromorphic electronics, will depend on their scalable synthesis in thin-film form with desired crystal quality, defect density, and phase purity.
Graphical abstract
Introduction
Neuromorphic computing broadly represents the use of non-von Neumann architectures to emulate learning exhibited by the human brain. The term “von Neumann architecture” represents any stored-program computer in which the fetch instruction and data operation may not occur simultaneously since they share a common bus, leading to the “von Neumann bottleneck” that describes the energy- and time-intensive transfer of data between separate memory and compute blocks. At advanced complementary metal oxide semiconductor (CMOS) nodes, the cost of accessing data from memory is two orders higher than multiplying two 32-bit numbers in processing units.1 Such a bottleneck restricts the ability of computing systems to execute data-intensive tasks whose demands are only growing with the advent of modern machine-learning models. Furthermore, a recent report shows that highly complex neural networks operating in the “overparameterized regime” do not overfit to spurious trends in training data but rather exhibit better generalization to unseen data than their less complex counterparts, motivating exponential growth each year in the number of model parameters since 2015 and the size of training datasets since 1988.2,3,4 In particular, the past decade has witnessed models from ResNet-50 ( model parameters) to Generative Pretrained Transformer 3 (GPT-3) ( model parameters) and datasets from ImageNet ( images) to JFT-3B ( images). By overcoming this bottleneck in electronic communication, clocking, thermal management, and power delivery, neuromorphic systems bring the promise of scalable hardware that can keep pace with the exponential growth of deep neural networks, leading us to define the first major thrust of neuromorphic computing: acceleration.2 Those neuromorphic systems concerned with acceleration are built for heightened speed and energy efficiency in existing machine-learning models and tend to have a relatively immediate impact. One common example would be crossbar arrays for vector-matrix multiplication (VMM) in the forward pass of deep neural networks. In contrast, we define the second major thrust of neuromorphic computing to be actualization, which is the realization of human neurobiological functions in non-von Neumann architectures. The second thrust will have a more delayed impact than the first but represents the hardware implementation for next-generation machine-learning models, with headway being made in the space of spiking neural networks (SNNs), Hebbian learning, and Hodgkin-Huxley neuron models.
Both acceleration and actualization approaches to neuromorphic computing require the development of specialized hardware. Memory technologies that support reading and writing can be historically categorized as either random-access memory (RAM) or flash memory. RAM is volatile, meaning it requires a power supply to refresh its memory states every few milliseconds, leading to undesirably large power consumption for neuromorphic applications.5 While Si-based flash memory represents the dominant non-volatile memory (NVM) device in the electronics industry, it also suffers from deficiencies with respect to neuromorphic systems, including slow programming speed, relatively high operating voltage, and poor electrical endurance.6 Furthermore, both RAM and flash memory are based on charge storage, which can use either a capacitor, crosscoupled inverter, or floating-gate transistor, and are currently struggling with scalability below 10-nm lateral dimensions. Both memories are also separate from the compute block, meaning they do not offer a solution to the von Neumann bottleneck. A new class of memory, often referred to as “emerging memory devices,” has been gaining much traction over the past decade due to their combination of fast switching speed in static RAM (SRAM), large storage density in dynamic RAM (DRAM), and non-volatility in flash memory, as well as their multi-bit/analog in-memory computing abilities.
Emerging NVM devices can be categorized into four different groups: resistive, phase change, ferroelectric, and ferromagnetic. We will also touch on a fifth group that combines the elements of the latter four in an optoelectronic setting, i.e., light pulse-induced programming and erasing of memory states. Resistive devices (resistive RAM [ReRAM]) exhibit resistive switching, which is a cyclical change of device resistivity via the formation/rupture of a conductive filament across distinct stable levels as induced by an electrical bias (Figure 1Ai).7 For a two-state resistive device, the low-resistance state (LRS) is achieved applying a positive voltage , while the high-resistance state (HRS) is achieved by applying a negative voltage (Figure 1Bi). Phase-change memory (PCM) is based on electrical contrast between highly resistive amorphous states and highly conductive crystalline states in phase-change materials (Figure 1Aii).8 In order to change the state of a phase-change device, high-power heat pulses are required to melt/quench the material between crystalline and amorphous states (Figure 1Bii). Ferroelectric devices (ferroelectric RAM [FeRAM]) leverage the change in current or resistance due to the induced polarization of a ferroelectric material with respect to a voltage bias (Figure 1Aiii).9 A two-state ferroelectric device requires a voltage to induce a remnant polarization in the ferroelectric material and a voltage to induce a remnant polarization in the ferroelectric material (Figure 1Biii). Ferromagnetic devices contain two ferromagnetic layers: a reference layer that is pinned to a fixed polarization and a free layer that can take on a programmed polarization (Figure 1Aiv).9 These two layers can either be parallel or anti-parallel, which induces an LRS or HRS, respectively (Figure 1Biv).
A brief comparison between these four memory technologies is warranted. ReRAM and PCM are known for their high bit density, fast read and write speeds, and lengthier endurance compared with other NVM technologies. However, their current-driven programming mechanism leads to a relatively high energy consumption, which can be a limiting factor in some applications. Magnetoresistive RAM (MRAM) has high endurance, low power consumption, and high speed, but achieving an ON/OFF ratio higher than 10 can be difficult. While ferroelectric memory offers a fast and low-power programming speed thanks to its field-driven programming mechanism, achieving both high retention and high bit density in FeRAM can be challenging.
This perspective focuses on the promising subset of emerging NVM devices for which 2D crystals form a critical component, as well as those devices that are entirely made from 2D crystals and their van der Waals heterostructures. Following the advent of stable monocrystalline graphitic films at the thickness of only a few atoms, the search for more stable 2D materials has yielded a new generation of emerging memory devices that offer even higher switching speed ( ns) and lower power loss ( pJ) while maintaining a low threshold voltage ( V) and lengthy retention time ( years).6,10 The additional knob of gate tunability enabled by 2D materials-based active layers is another exciting advantage for neuromorphic applications, as it gives rise to a host of multi-state/analog memory devices. Further, 2D materials open the door to non-volatile optoelectronic memory due to their strong light-matter interactions and significant photogenerated charge trapping.11 The goal of this perspective will be to connect how 2D materials engender better memory devices that in turn create architectures and algorithms for next-generation machine-learning models. We will then hypothesize about the future direction of neuromorphic computing and illustrate the chasm that must be bridged between current machine-learning hardware and the human brain.
Section snippets
Past
The human brain is known for its high level of energy efficiency and low latency in processing complex information. It is able to perform better than any computer, especially for tasks that require processing large amounts of complex information, such as perception. These benefits are largely due to the brain’s dense network of neurons that transmit signals efficiently through synapses. Some researchers saw the potential for electronic technology to model the human brain system. For example,
Present
The performance of memory cells is a crucial factor for in-memory computing systems. In this section, we will review the most recent developments in 2D memories for in-memory neuromorphic computing technology, including device performance, design considerations, and potential applications. 2D memories can be classified into four main categories based on their working mechanisms: resistive devices, phase-change devices, ferroelectric devices, and ferromagnetic devices. As alluded to previously,
Problems
The above sections describe major recent developments in 2D materials-based NVM memory devices for application in neuromorphic computing architectures. While significant progress has been achieved, there are several major technological challenges that largely reside at the materials level. Though demonstrations at the individual-device level remain impressive, any complex microelectronic system requires more devices on the order of several magnitudes.67 To move from device to architecture
Possibilities
The key benefit of 2D materials in NVM and neuromorphic devices lies in just one major aspect of their physical property: atomically small thickness. This property leads to two performance advantages in most devices: voltage scaling and gate tunability. Voltage scaling, while very important and necessary, will inherently compromise endurance and reliability for ReRAM devices. Therefore, 2D materials-based ReRAM devices are unlikely to surpass oxide-based ReRAM for the foreseeable future.
Acknowledgments
D.J. and X.L. acknowledge support from Intel RSA. K.K. acknowledges support from the National Science Foundation (NSF) Graduate Research Fellowship Program (GRFP), fellow ID: 2022338725.
Author contributions
All authors conceived the structure and contents of the manuscript. K.K. and X.L. did the majority of the writing and literature research under the supervision of D.J. All authors read, commented on, and edited the manuscript.
Declaration of interests
The authors declare no competing interests.
References (100)
- et al.
Analogue signal and image processing with large memristor crossbars
Nat. Electron.
(2017) - et al.
Ionic modulation and ionic coupling effects in MoS2 devices for neuromorphic computing
Nat. Mater.
(2019) - et al.
Graphene–ferroelectric transistors as complementary synapses for supervised learning in spiking neural network
npj 2D Mater. Appl.
(2019) - et al.
Retinomorphic optoelectronic devices for intelligent machine vision
iScience
(2022) - et al.
Monatomic 2D phase-change memory for precise neuromorphic computing
Appl. Mater. Today
(2020) - et al.
Ionic modulation and ionic coupling effects in MoS(2) devices for neuromorphic computing
Nat. Mater.
(2019) 1.1 Computing's energy problem (and what we can do about it)
- et al.
Freely scalable and reconfigurable optical hardware for deep learning
Sci. Rep.
(2021) - et al.
Trends in Training Dataset Sizes
- et al.
Reconciling modern machine-learning practice and the classical bias–variance trade-off
Proc. Natl. Acad. Sci. USA
(2019)
Resistive random access memory (RRAM): an overview of materials, switching mechanism, performance, multilevel cell (mlc) storage, modeling, and applications
Nanoscale Res. Lett.
Decade of 2D-materials-based RRAM devices: a review
Sci. Technol. Adv. Mater.
Recommended methods to study resistive switching devices
Adv. Electron. Mater.
Phase change memory technology
J. Vac. Sci. Technol. B Nanotechnol. Microelectron.
In-memory computing with resistive switching devices
Nat. Electron.
Electric field effect in atomically thin carbon films
Science
Two-terminal multibit optical memory via van der Waals heterostructure
Adv. Mater.
Currents carried by sodium and potassium ions through the membrane of the giant axon of Loligo
J. Physiol.
Neuromorphic systems: past, present and future
Mathematical models of threshold phenomena in the nerve membrane
Bull. Math. Biophys.
An active pulse transmission line simulating nerve axon
Proc. IRE
Analog VLSI Implementation of Neural Systems
TrueNorth: design and tool flow of a 65 mW 1 million neuron programmable neurosynaptic chip
IEEE Trans. Comput. -Aided. Des. Integr. Circuits Syst.
Efficient reward-based structural plasticity on a SpiNNaker 2 prototype
IEEE Trans. Biomed. Circuits Syst.
A hardware-deployable neuromorphic solution for encoding and classification of electronic nose data
Sensors
Loihi: a neuromorphic manycore processor with on-chip learning
IEEE Micro
Pathways to efficient neuromorphic computing with non-volatile memory technologies
Appl. Phys. Rev.
Memory devices and applications for in-memory computing
Nat. Nanotechnol.
Neuro-inspired computing with emerging nonvolatile memorys
Proc. IEEE
NeuroSim: a circuit-level macro model for benchmarking neuro-inspired architectures in online learning
IEEE Trans. Comput. -Aided. Des. Integr. Circuits Syst.
Face classification using electronic synapses
Nat. Commun.
Equivalent-accuracy accelerated neural-network training using analogue memory
Nature
Fully memristive neural networks for pattern classification with unsupervised learning
Nat. Electron.
Metal–oxide RRAM
Proc. IEEE
Nonvolatile memories based on graphene and related 2D materials
Adv. Mater.
Emerging 2D memory devices for in-memory computing
Adv. Mater.
Memory applications from 2D materials
Appl. Phys. Rev.
Atomristor: nonvolatile resistance switching in atomic sheets of transition metal dichalcogenides
Nano Lett.
Robust memristors based on layered two-dimensional materials
Nat. Electron.
All WSe(2) 1T1R resistive RAM cell for future monolithic 3D embedded memory integration
Nat. Commun.
Wafer-scale integration of two-dimensional materials in high-density memristive crossbar arrays for artificial neural networks
Nat. Electron.
Wafer-scale 2D hafnium diselenide based memristor crossbar array for energy-efficient neural network hardware
Adv. Mater.
Energy-efficient phase-change memory with graphene as a thermal barrier
Nano Lett.
Phase transitions in 2D materials
Nat. Rev. Mater.
Electric-field induced structural transition in vertical MoTe2- and Mo1–xWxTe2-based resistive memories
Nat. Mater.
Electrically driven reversible phase changes in layered In2Se3 crystalline film
Adv. Mater.
Post-CMOS compatible aluminum scandium nitride/2D channel ferroelectric field-effect-transistor memory
Nano Lett.
A ferroelectric semiconductor field-effect transistor
Nat. Electron.
Room-temperature ferroelectricity in CuInP2S6 ultrathin flakes
Nat. Commun.
An electronic synapse based on 2D ferroelectric CuInP2S6
Adv. Electron. Mater.
Cited by (2)
- 2
These authors contributed equally