摘要
在信息爆炸的时代,解决计算中不断升级的能源需求变得至关重要.传统的计算机系统面临性能限制,促使人们寻找新的计算范式.自旋电子器件以其非易失、速度快、高能效等特点,有望解决传统电子器件所面临的困境.本文从自旋人工电子突触、自旋振荡器神经元、自旋人工神经网络和概率神经网络四个方面综述了基于自旋的智能器件和物理神经网络的最新研究进展.自旋电子器件的蓬勃发展及其在智能计算和信息存储等领域展现出的应用潜力有望激发新一轮的信息技术革命.
In the era of information explosion, addressing computational bottlenecks and escalating energy demands has becomecrucial. Conventional computer systems face performance limitations, prompting the search for novel computationalparadigms. Spin-based electronic devices, renowned for their non-volatility, fast response and high energy efficiency, offerpromising solutions. This review outlines recent progress in spin-based intelligent devices and physical neural networks,focusing on spintronic synapses, spin oscillator neurons, spin-based artificial neural networks (ANNs), and probabilisticneural networks.Spin-orbit torque (SOT) devices are emerging as attractive candidates for artificial synapses due to their non-volatilityand tunability. By exploiting methods like interlayer exchange coupling, exchange bias, and magnetic field modulation,researchers have achieved stable multi-state switching in SOT devices, mimicking synaptic weight changes. This enablesfunctionalities like spike-timing-dependent plasticity (STDP), critical for learning and information storage. Devicesutilizing topological magnetic structures like skyrmions show linear weight updates, enhancing performance.Inspired by the nonlinear dynamics of biological neurons, spin-torque nano-oscillators (STNOs) and spin-Hall nanooscillators(SHNOs) exhibit similar behavior. STNOs utilize spin-transfer torque, while SHNOs leverage spin Hall effectfor microwave oscillations. Applications in recurrent neural networks (RNNs) for speech recognition and radiofrequency(RF) signal identification have demonstrated their potential in neuromorphic computing.Building on the low power consumption and non-volatility of spin devices, spin-based ANNs have emerged as viablealternatives to CMOS-based networks. Arrays of spintronic devices facilitate in-memory computing, significantly reducingenergy consumption. Studies using STT-MRAM and SOT devices have achieved binary and analog weight neural networkcomputations, demonstrating their suitability for edge computing and IoT applications.Recognizing the role of stochasticity in biological neural networks, spintronic devices with low energy barriers exhibitintrinsic randomness, enabling probabilistic computation. By manipulating the switching probability of magnetic tunneljunctions (MTJs), probabilistic bits mimic neurons capable of solving complex problems like Ising model optimization.High-energy barrier SOT devices retain non-volatility while allowing tunable stochasticity, advancing the field.Recent advancements in spin-based devices have paved the way for the development of physical neural networks thatmimic the functionalities of biological neural systems. By precisely controlling magnetic moments, researchers aredesigning circuits and chips that can dynamically adjust their connections, mirroring the cognitive and learning abilities ofthe brain. This groundbreaking technology promises profound impacts on neuromorphic computing, brain-inspired AI, andaddressing the bottlenecks in current AI systems related to efficiency, energy consumption, and adaptability.
作者
刘祥语
文辉
雷坤
兰修凯
王开友
Xiangyu Liu;Hui Wen;Kun Lei;Xiukai Lan;Kaiyou Wang(State Key Laboratory for Superlattices and Microstructures,Institute of Semiconductors,Chinese Academy of Sciences,Beijing 100083,China;College of Materials Science and Opto-Electronic Technology,University of Chinese Academy of Sciences,Beijing 100049,China)
出处
《科学通报》
北大核心
2025年第13期1874-1883,共10页
Chinese Science Bulletin
基金
国家重点研发计划(2022YFA1405100)
国家自然科学基金(12241405,12204473,12274406,12104449)资助。
关键词
自旋电子器件
物理神经网络
自旋轨道矩
存算一体
神经形态计算
spintronic devices
physical neural networks
spin-orbit torque
in-memory computing
neuromorphic computing
