In‐Memory Mathematical Operations with Spin‐Orbit Torque Devices

Abstract

Analogue arithmetic operations are the most fundamental mathematical operations used in image and signal processing as well as artificial intelligence (AI). In-memory computing offers high performance and energy-efficient computing paradigm. To date, in-memory analogue arithmetic operation with emerging nonvolatile devices were usually implemented using discrete components, which limits the scalability and blocks large scale integration. Here, we experimentally demonstrate a prototypical implementation of in-memory analogue arithmetic operations (summation, subtraction and multiplication), based on in-memory electrical current sensing units using spinorbit torque (SOT) devices. The proposed analogue arithmetic operation structures are smaller than the state-of-the-art CMOS counterparts by several orders of magnitude.Moreover, data to be processed and computing results can be locally stored, or the analogue computing can be done in the nonvolatile SOT devices, which were exploited to experimentally implement image edge detection and signal amplitude modulation with simple structure. Furthermore, we constructed an artificial neural network (ANN) with SOT devices based synapses to realize pattern recognition with high accuracy of ~95%.

Memristor-Based Artificial Chips

“…42 On the other hand, because the volume of a memristor can be reduced to the nanoscale, even to an atomic level, it can show quantum conduction and quantum switching phenomena, thus providing a possibility for the development of quantum computers. 43 Therefore, the research of memristor-based brain-like chip integration has very important practical significance and application prospects.…”

“…At the same time, because the memristor has a typical two-teminal crossbar configuration, it is compatible with many existing CMOS device structures and manufacturing processes, and so it provides a good foundation for further chip integration . On the other hand, because the volume of a memristor can be reduced to the nanoscale, even to an atomic level, it can show quantum conduction and quantum switching phenomena, thus providing a possibility for the development of quantum computers . Therefore, the research of memristor-based brain-like chip integration has very important practical significance and application prospects.…”

Restricted Boltzmann Machines Implemented by Spin–Orbit Torque Magnetic Tunnel Junctions

“…Especially, nonvolatile spintronic devices like magnetic tunnel junctions (MTJs) with high speed and endurance could potentially be a game-changer − despite their rareness in this field before. The intrinsically stochastic spin dynamics of MTJs, − especially those employing SOT, − make them competitive true random number generators (TRNGs), − thereby extending their utility to generative AI applications. Here, we aim to address the stochastic sampling challenge in spintronic hardware, enhance the efficiency of the Gibbs sampling operation in RBM by leveraging high-performance SOT-MTJs, and experimentally demonstrate the high compatibility of SOT-MTJs with the needs of RBM.…”