PaperlayerView exact revision on arXiv

Source-linked AI summary

Spintronics for neuromorphic computing

J. Grollier, D. Querlioz, K. Y. Camsari, K. Everschor-Sitte, S. Fukami, M. D. Stiles

arXiv:2007.06092v1physics.app-phcond-mat.mtrl-sci

TL;DR

Traditional neuromorphic circuits face energy and area constraints, motivating spintronic alternatives. This review examines spintronic devices for neuromorphic computing and reports proof-of-concept results in associative memory and neural-network tasks.

  • Problem

    Traditional electronic neurons and synapses remain limited by their energy and area requirements, while analog, plastic, non-volatile synaptic storage remains needed.

  • Method

    The review synthesizes spintronic implementations using magnetic tunnel junctions, magnetic textures, memristors, and spin-torque nano-oscillators as neuromorphic elements.

  • Results

    Spintronic systems achieved 91.2% lower associative-memory power, up to 99.6% spoken-digit precision, and 89% vowel-recognition accuracy after training.

  • Takeaways & Limitations

    Early experiments provide proof of concept for locally combining computation and memory and exploiting multiphysics in brain-inspired spintronic systems.

  • Takeaways & Limitations

    Scaling spintronic neuromorphic systems remains challenging because magnetic tunnel junctions are difficult to read rapidly and precisely, especially in multistate operation.

Abstract

from arXiv · show

Neuromorphic computing uses brain-inspired principles to design circuits that can perform computational tasks with superior power efficiency to conventional computers. Approaches that use traditional electronic devices to create artificial neurons and synapses are, however, currently limited by the energy and area requirements of these components. Spintronic nanodevices, which exploit both the magnetic and electrical properties of electrons, can increase the energy efficiency and decrease the area of these circuits, and magnetic tunnel junctions are of particular interest as neuromorphic computing elements because they are compatible with standard integrated circuits and can support multiple functionalities. Here we review the development of spintronic devices for neuromorphic computing. We examine how magnetic tunnel junctions can serve as synapses and neurons, and how magnetic textures, such as domain walls and skyrmions, can function as neurons. We also explore spintronics-based implementations of neuromorphic computing tasks, such as pattern recognition in an associative memory, and discuss the challenges that exist in scaling up these systems.

II-Spintronic synapses · b. Exploiting the inherent stochastic switching in binary magnetic tunnel

Spintronic synapses embed magnetic tunnel junction memory within neural-network hardware, reducing memory-transfer costs. Their inherent switching errors can be tolerated by neural networks and may provide probabilistic weight updates during training.

  • a. Embedding memory in the processor: Magnetic tunnel junction memory cells store synaptic weights in associative-memory hardware neural networks.This approach embeds synaptic storage in the neural-network implementation.
  • a. Embedding memory in the processor: 13.6× lower memory needs and 89% lower energy consumption were achieved than with a non-neural content-addressable-memory search architecture.The comparison concerns Jarollahi et al.’s magnetic tunnel junction-based logic-in-memory content-driven search engine.
  • a. Embedding memory in the processor: A 4T-2MTJ spin-transfer torque magnetoresistive random-access memory associative processor autonomously activates only currently accessed memory cells.The intelligent powering strategy drastically reduced energy consumption.
  • a. Embedding memory in the processor: 91.2% lower power consumption was obtained than with a twin chip using six-transistor static random-access memory.This reduction resulted from the intelligent powering strategy in the associative processor.
  • b. Exploiting the inherent stochastic switching in binary magnetic tunnel: Thermal activation makes magnetic tunnel junction switching inherently prone to bit errors, which conventional applications address through higher energy barriers, error-correcting codes, or specialized write strategies.Higher energy barriers lead to higher programming currents.
  • b. Exploiting the inherent stochastic switching in binary magnetic tunnel: Neural networks can tolerate relatively high synaptic error rates because they are inherently resilient to bit errors.Thus, magnetic tunnel junction synapses need not meet the reliability required for conventional computing.
  • b. Exploiting the inherent stochastic switching in binary magnetic tunnel: Training can exploit programming errors by applying larger synaptic-weight changes with reduced probabilities instead of repeated small adjustments.Magnetic tunnel junctions can implement probabilities in a regime with high bit—although the supplied passage ends before completing the condition.

c. Spintronic memristors

Spintronic memristors address the mismatch between binary magnetic tunnel junctions and real-valued synaptic weights by storing analog information in magnetic or antiferromagnetic states. Their nonvolatility and high endurance support learning, memory, and many learning cycles in neuromorphic synapses.

  • c. Spintronic memristors: Real-valued synaptic weights make binary magnetic tunnel junctions inefficient because many junctions are needed to store one weight.Binary magnetic information is encoded by magnetization pointing up or down.
  • c. Spintronic memristors: Memristors are analog, nonvolatile, and plastic resistors particularly suited to imitating synapses.Their development includes early hardware-synapse use, theoretical circuit-element treatment, and experimental Pt-TiO2-x-Pt nanodevices.
  • c. Spintronic memristors: Spintronic memristors store analog information through domain-wall position, skyrmion number, or current-controlled Néel-vector states.Domain-wall displacement produces lower or higher resistance depending on position; skyrmion number represents analog information; antiferromagnetic devices use current-induced Néel-vector control.
  • c. Spintronic memristors: Nonvolatility enables simultaneous learning and memory, while high endurance permits an outstanding number of learning cycles.These properties make spintronic memristors building blocks for neuromorphic computing with artificial synapses and are important for adaptive applications.

III-Spintronic neurons · a. Spin-torque nano-oscillators

Spin-torque nano-oscillators are magnetic tunnel junctions with memory, nonlinear oscillations and tunability that support neuromorphic neuron functions. A single oscillator has also demonstrated time-multiplexed emulation of a 400-neuron reservoir computer.

  • III-Spintronic neurons: Neurons have more complicated features than the simple nonlinear functions commonly targeted in nanoscale hardware neural networks.
  • III-Spintronic neurons: High biological recording noise can make spike trains seemingly random, motivating Poisson and stochastic models for neuromorphic computing.
  • a. Spin-torque nano-oscillators: Spin-torque nano-oscillators are magnetic tunnel junctions driven into spontaneous microwave oscillations by injected direct current.
  • a. Spin-torque nano-oscillators: Their oscillation amplitudes retain memory through finite magnetization relaxation, enabling imitation of leaky integration in neurons.
  • a. Spin-torque nano-oscillators: Stable, persistent behavior, nonlinear frequency and amplitude responses, and high tunability make these oscillators suitable for activation functions and synchronization.
  • a. Spin-torque nano-oscillators: A single spin-torque nano-oscillator emulated a full neural network of 400 neurons through time-multiplexing.The oscillator periodically assigned time intervals to individual neuron states and used finite relaxation time to emulate coupling.

b. Superparamagnetic tunnel junctions

Superparamagnetic tunnel junctions exploit stochastic switching for low-energy neuromorphic computation, supporting random-bit generation, Poisson-like signaling, and controllable switching rates. Their average states also enable stochastic computing between deterministic bits and qubits, using unstable nanomagnets for ultra-low-power implementations.

  • b. Superparamagnetic tunnel junctions: Brain-inspired neuromorphic designs can exploit the stochastic effects of magnetic tunnel junctions because biological synapses and neurons are partly stochastic.Brain processes may trade reliability for energy efficiency, and biological neurons are sometimes modeled as Poisson neurons.
  • b. Superparamagnetic tunnel junctions: Superparamagnetic tunnel junctions generate random bits by directly reading their state, an extremely low-energy operation.Their switching rate can be controlled through spin torques and magnetic fields.
  • b. Superparamagnetic tunnel junctions: Superparamagnetic tunnel junctions resemble Poisson neurons, although their outputs are telegraph signals rather than spike trains.The junction switching rate is tunable through spin torques and magnetic fields.
  • b. Superparamagnetic tunnel junctions: Using the average junction state provides a stochastic computing regime between deterministic digital bits and quantum-computing qubits.This approach replaces stable free layers with unstable nanomagnets by reducing anisotropy or total magnetic moment.
  • b. Superparamagnetic tunnel junctions: Magnetic-tunnel-junction implementations of p-bits may enable ultra-low-power stochastic computing reminiscent of brain processes.P-bits can also be implemented using CMOS circuits.

c. Domain-wall and skyrmion based neurons · IV-Neuromorphic computing with small spintronic systems

Domain walls and skyrmions are magnetic-soliton information vectors that can be manipulated by spin torques and may support unconventional computing through nonlinear resistance. System-level spintronic neuromorphics can propagate information between devices but must address coupling control and device variability.

  • c. Domain-wall and skyrmion based neurons: Magnetic solitons, including domain walls and skyrmions, are proposed as alternative neuron types.These objects can serve as vectors of information for computing.
  • c. Domain-wall and skyrmion based neurons: Spin torques and spin-orbit torques can manipulate and move domain walls and skyrmions over large distances.
  • c. Domain-wall and skyrmion based neurons: Nonlinear resistance changes in magnetic skyrmion systems can be exploited for unconventional computing.The changes arise from magnetoresistance effects combined with spin-(orbit)-torques that move or distort skyrmions.
  • IV-Neuromorphic computing with small spintronic systems: Spintronic neuromorphic systems can harness spin currents, spin waves, or microwave emissions to propagate information between devices.
  • IV-Neuromorphic computing with small spintronic systems: Assembling spintronic neurons and synapses directly into systems requires controlling their coupling.
  • IV-Neuromorphic computing with small spintronic systems: System-level spintronic neuromorphics must also address inevitable device variability.Research has begun addressing coupling control and variability challenges.

a. Computing with spintronic memristors

Spintronic memristors can act as trainable artificial synapses, enabling networks to store information through tuned synaptic states. A proof-of-concept Hopfield associative memory used spin-orbit-torque memristive devices to associate 3×3 block patterns.

  • Learning with spintronic memristors: Spintronic memristors function as artificial synapses whose states are tuned during training to store information collectively.Updating device states in response to new information provides the learning capability required of artificial synapses.
  • Associative memory: Borders et al. demonstrated a proof-of-concept associative memory using an artificial neural network with spintronic synapses.The implementation used antiferromagnet/ferromagnet spin-orbit-torque switching devices with memristive functionality.
  • Associative memory: The Hopfield model memorizes and associates patterns through a variable synaptic-weight matrix encoding stored information.Each neuron is connected to all other neurons via synapses with variable weights.
  • Associative memory: The experiment associated three kinds of 3×3 block patterns in 9-neuron systems using 36 spin-orbit-torque-based memristive devices.The 36 devices correspond to the symmetry of the synaptic weight matrix, while field-programmable gate arrays emulated neurons.

b. Computing with synchronized spin-torque nano-oscillators · c. Computing with superparamagnetic magnetic tunnel junctions

Synchronized spin-torque nano-oscillators enabled trained neural-network classification, while superparamagnetic magnetic tunnel junctions supported probabilistic spiking, energy-efficient learning, and p-bit computation. These approaches demonstrate neuromorphic and optimization functions but require scalable oscillator arrays, extended learning rules, and careful device-level design.

  • b. Computing with synchronized spin-torque nano-oscillators: Four coupled spin-torque nano-oscillators classified seven American vowels after fewer than hundred training iterations.The network used synchronization among oscillators to perform signal classification at microwave frequencies.
  • b. Computing with synchronized spin-torque nano-oscillators: 89 % recognition on test data, or 84 % after cross validation, exceeded a similarly parameterized multilayer perceptron.The reported classification performance followed training on the vowel-recognition task.
  • b. Computing with synchronized spin-torque nano-oscillators: Scaling synchronization-based networks requires hundreds of oscillators with different frequencies but similar synchronization ranges.The oscillator arrays must preserve compatible synchronization behavior across devices.
  • b. Computing with synchronized spin-torque nano-oscillators: The demonstrated simple learning rule may not extend readily to deep networks, making tunable oscillator coupling important for multilayer systems.The proposed extension changes coupling between oscillators rather than their individual frequencies.
  • c. Computing with superparamagnetic magnetic tunnel junctions: Temperature-driven fluctuations in superparamagnetic magnetic tunnel junctions can imitate neural Poisson spiking dynamics.Under electrical current and spin torque, their mean frequency shows a bell-shaped response.
  • c. Computing with superparamagnetic magnetic tunnel junctions: 23 nJ per operation during learning and 7.4 nJ after learning were consumed by a 128-input, 128-output CMOS-spintronic system, versus 330 nJ for low-power spiking CMOS neurons.The reported circuit combined CMOS electronics with spintronic junctions for the application.
  • c. Computing with superparamagnetic magnetic tunnel junctions: Superparamagnetic tunnel junctions with EB≈ kBT can form p-bit circuits that stochastically search hard-problem solution spaces at megahertz-to-gigahertz speeds.These networks operate massively in parallel and asynchronously, encoding solutions as low-energy states.
  • c. Computing with superparamagnetic magnetic tunnel junctions: Classical p-circuits can guide networks toward energy minima and implement invertible logic, allowing multiplier circuits to operate in reverse for number factorization.This reverse operation follows from the reciprocal nature of p-circuits.

d. Computing with nanomagnets

Nanomagnets can support computation through direct dipolar coupling and energy minimization, reducing CMOS overhead in spintronic circuits. Nanomagnet arrays have been demonstrated for solving Ising Hamiltonians.

  • d. Computing with nanomagnets: Direct dipolar coupling between nanomagnets enables computation through energy minimization while reducing CMOS overhead.This approach avoids relying exclusively on CMOS circuits or resistive crossbar arrays to couple junctions.
  • d. Computing with nanomagnets: Several demonstrations have used nanomagnet arrays to solve Ising Hamiltonians.These demonstrations exploit the physical dipolar interactions among nanomagnets.

e. Computing with skyrmions · V-Challenges for scaling up

Skyrmion assemblies provide a simulated reservoir-computing fabric with history-dependent, nonlinear responses that enable simple pattern classification. Scaling spintronic neuromorphic systems toward useful pattern recognition remains challenging because software networks already contain hundreds of millions of interconnected neurons and synapses.

  • e. Computing with skyrmions: Skyrmion assemblies in conducting thin films are proposed as a reservoir-computing fabric for neuromorphic systems.Input signals are injected through voltage patterns, and outputs are obtained from resistance measurements between contacts.
  • e. Computing with skyrmions: Pinned skyrmions exhibit nonlinear I-V characteristics because spin torques deform them, changing current patterns and measured resistance.For a single skyrmion, the nonlinearity is small because it couples only to deformation size; larger effects favor skyrmion assemblies.
  • e. Computing with skyrmions: Voltage-pattern responses depend on reservoir history, demonstrating short-term memory in the simulated skyrmion system.Its complex magnetic response patterns provide high-dimensional nonlinear filtering of input signals.
  • e. Computing with skyrmions: The simulated skyrmion reservoir shows its strongest nonlinearity near the natural time scale of ferromagnetic systems, in the nanoseconds.Simulations demonstrate simple pattern classification.
  • V-Challenges for scaling up: Early experimental demonstrations highlight spintronics’ promise for future neuromorphic-computing applications.These demonstrations involve small spintronics systems.
  • V-Challenges for scaling up: Scaling spintronic systems to useful pattern-recognition sizes requires overcoming challenges specific to spintronics and challenges shared by other approaches.The section frames scaling as a major hurdle following the first experimental demonstrations.
  • V-Challenges for scaling up: Software deep networks for image recognition already comprise hundred millions of interconnected neurons and synapses, setting the scale required for useful pattern recognition.Spintronic systems must be scaled toward comparable usefulness, although the passage does not specify a target hardware size.

a. Adapting algorithms to spintronic hardware · b. Low energy

Spintronic neuromorphic hardware must adapt algorithms and readout methods to device nonidealities, while reducing CMOS overhead and device-level energy consumption. Key constraints include weak resistance contrasts, nonlinear and asymmetric weight updates, and higher write than read energy in magnetic tunnel junctions.

  • a. Adapting algorithms to spintronic hardware: Magnetic tunnel junctions are difficult to read quickly because their resistance changes are small, especially in multistate memristive-like operation.Their typical OFF/ON ratios are between one and three.
  • a. Adapting algorithms to spintronic hardware: Multistate magnetic tunnel junctions have typically lower OFF/ON ratios than other resistive switching cells.Other resistive switching cells have ratios ranging from tens to millions.
  • a. Adapting algorithms to spintronic hardware: On-chip neural-network training imposes additional constraints because backpropagation based on gradient descent requires highly linear and symmetric weight variations.This requirement is an issue for emerging memories and most memristor types.
  • a. Adapting algorithms to spintronic hardware: Most memristor types exhibit highly nonlinear and asymmetric responses, complicating direct implementation of gradient-descent training.The passage identifies this behavior as an issue for emerging memories and most memristor types.
  • b. Low energy: Spintronic neuromorphic systems should reduce CMOS overhead by exploiting physical effects to perform functions that CMOS does not do well.The passage frames this as a system-level energy-reduction priority.
  • b. Low energy: Low energy consumption is also important for individual spintronic devices, including magnetic tunnel junctions.As in most nonvolatile memories, magnetic tunnel junction write energy is higher than read energy.

c. Interconnection · Data availability

Spintronics addresses neuromorphic hardware’s interconnection challenge by supporting dense, low-power connectivity while combining computation and memory locally and exploiting multiphysics. The study’s datasets are available from the corresponding authors on reasonable request.

  • c. Interconnection: Neuromorphic hardware must increase interconnection between neurons while controlling associated power consumption.Current algorithms use 10–1000 synapses per neuron, compared with 10,000 in the cortex.
  • c. Interconnection: Current approaches lack a good solution for achieving cortical-scale interconnection without high power consumption.The passage identifies this as a major challenge for neuromorphic hardware.
  • c. Interconnection: Spintronics offers opportunities to address neuromorphic interconnection through multilayer systems that naturally stack.The supplied passage introduces multilayer stacking as a spintronic opportunity in this domain.
  • c. Interconnection: Spintronics can enmesh computation and memory at a very local level.This is one of at least two ways spintronics could help realize artificial intelligence based on brain-computation principles.
  • c. Interconnection: Spintronics can exploit rich multiphysics as a source of computational power.The passage presents multiphysics exploitation as the second identified route toward artificial intelligence.
  • c. Interconnection: Recent experimental progress provides initial proofs of concept for brain-inspired spintronic systems.These results push toward development of large-scale systems.
  • Data availability: The study’s generated and analysed datasets are available from the corresponding authors on reasonable request.This statement applies to the study’s data availability.
Loading 2007.06092v1…