Technical Review of spin-based computing

TL;DR

The paper addresses the pressure on CMOS technologies from power and scaling limits and surveys spin-based computing as a promising alternative. It synthesizes four main approaches—RF spintronic synapses/neurons, spintronic p-bits, magnetic reservoir computing, and magnetic Ising machines—detailing their working principles, hardware metrics, and task-specific benchmarking frameworks. By distinguishing task-independent metrics (nonlinearity NL, memory MC, and IPC) from task-dependent performance (e.g., NARMA benchmarks, TTS), the work provides a structured foundation for fair cross-technology comparisons and identifies key challenges in training, co-design, and CMOS integration. The authors argue for continued material and device innovations, along with hybrid architectures and standardized benchmarking, to realize scalable, energy-efficient spin-based neuromorphic hardware with broad impact on AI, optimization, and real-time processing.

Abstract

Spin-based computing is emerging as a powerful approach for energy-efficient and high-performance solutions to future data processing hardware. Spintronic devices function by electrically manipulating the collective dynamics of the electron spin, that is inherently non-volatile, nonlinear and fast-operating, and can couple to other degrees of freedom such as photonic and phononic systems. This review explores key advances in integrating magnetic and spintronic elements into computational architectures, ranging from fundamental components like radio-frequency neurons/synapses and spintronic probabilistic-bits to broader frameworks such as reservoir computing and magnetic Ising machines. We discuss hardware-specific and task-dependent metrics to evaluate the computing performance of spin-based components and associate them with physical properties. Finally, we discuss challenges and future opportunities, highlighting the potential of spin-based computing in next-generation technologies.

Figures (5)

• Figure 1: Four spin-based computing technologies discussed in this article. Conceptual diagrams and proof-of-concepts for a RF spintronic neural network, b probabilistic computer, c reservoir computer, and d spin-based Ising machine. Details of each system are discussed in the main text.
• Figure 2: Types of stochastic magnetic tunnel junctions (MTJ) designs. a A fixed‐free MTJ with perpendicular magnetic anisotropy (PMA) exhibits bias dependence $[\ddagger]$10019530 and can reach nanosecond‐scale fluctuation speeds $[\ast]$soumah2024nanosecond. b A fixed‐free MTJ with in‐plane magnetic anisotropy (IMA) also shows bias dependence $[\S']$singh2023hardware (unpublished data) and demonstrates microsecond‐scale p‐bit fluctuations with a 50/50 probability $[\S]$singh2023hardware. c MTJs incorporating synthetic antiferromagnets (SAF) in both the fixed and free layers can be either bias‐dependent or bias‐independent, depending on the extent of IMA $[\dagger]$sun2023stochastic. Alternatively, using an SAF for the free layer offers robustness against external fields and sub‐microsecond switching $[+]$PhysRevApplied.18.054085. d Removing the fixed layer yields a double‐free MTJ, which is effectively bias‐independent, as evidenced by the fraction of time spent in antiparallel (AP) vs. parallel (P) states under any bias, with sub‐microsecond dynamics $[\P]$10.1063/5.0219606. e Theoretically, an SAF design for both free layers is proposed to be bias‐independent and free of dipolar coupling for any device diameter, exhibiting nanosecond time scales $[\bullet]$PhysRevApplied.21.054002.
• Figure 3: Reservoir computing metrics evaluations. a, Spatial analysis of NL for a computationally defined skyrmion reservoir with an electrical input. NL is distributed inhomogeneously due to the spatial distribution of magnetic skyrmions and their nonlinear dynamics excited by the current. b,Comparisons between the test (red solid line) and reconstructed (blue dotted line) data with different $k$ values. c, MC($k$) obtained by a single shot (red dotted) and averaged data (blue solid) as a function of $k$ in physical reservoirs based on a STNO. d, First-(top), second-(middle) and their-order(bottom) IPC of a physical reservoir based on spintronic oscillators, externally controlled by feedback gain. Panel a adapted with permission from Ref. love2023spatial, American Physical Society; Panels b-c adapted with permission from Ref. Tsunegi_JJAP2018, IOP Publishing; Panel d adapted with permission from ref. Tsunegi_AdvIntSys2023, Wiley-VCH GmbH.
• Figure 4: SWIM time traces for the measurement of time-to-saturation $\tau_{sat}$ and time-to-solution $\tau_{solution}$ parameters. a A control signal turning on/off the SWIM. b The propagating RF pulses colored according to their instantaneous phase. c Phase values in the center of each propagating RF pulse sampled at each circulation period. Spin switching scenarios for 4- and 8-spin MAX-CUT problem computations. Envelopes (d, e) and instantaneous phase signals (f, g) of the 3rd propagating RF pulse signal within 14 circulation periods plotted in the form of overlapping traces in a relative scale for 4-spin (d, f) and 8-spin (e, g) MAX-CUT optimization problem. The color of the time traces corresponds to the circulation number. All figures reproduced from Ref. litvinenko2023spinwave with permission.
• Figure 5: Magnetic Ising machine benchmarking. The table considers two distinct IM architectures -- spacial and time-multiplexed, parameters taken from corresponding references dutta_experimental_2019dutta2021isinghoushang2022phasealbertsson2021ultrafastlitvinenko2023spinwaveHonjo2021SciAdv100kCIM. For comparison, we also include the noisy mean-field GPU algorithm, phase-transition oscillator IMs, and Coherent Ising Machines (CIMs).