Spintronic Devices for Neural Networks

Representing the human brain in computers, so-called neuromorphic computing, is one of the focuses of interest in recent electronics as the brain is a model system that can readily accomplish complex tasks at small power consumption level, in contrast to conventional von Neumann computers. An artificial neural network offers a promising approach for low-power and intelligent neuromorphic computing. A key ingredient for the network is an artificial synapse that has analog and nonvolatile memory functionalities and frequent learning capability as in the real synapse in the brain. Here we show an artificial neural network with spintronic artificial synapses, which meet these requirements [1]. We will first describe an analog spin-orbit torque (SOT) switching device consisting of an antiferromagnet/ferromagnet bilayer structure [2], which can serve as the artificial synapse. The mechanism of analog and nonvolatile property [3] will be described. Secondly, a proof-of-concept demonstration of an artificial neural network with 36 SOT devices will be shown. An associative memory operation is performed based on the Hopfield model and learning ability of the spintronic artificial synapse is confirmed [4].
This work is jointly carried out with H. Ohno, W. A. Borders, A. Kurenkov, S. Sato, Y. Horio, and H. Akima of Tohoku University and P. Gambardella of ETH Zurich.
[1] S. Fukami and H. Ohno, J. Appl Phys. 124, 151904 (2018).
[2] S. Fukami et al., Nature Mater. 15, 535 (2016).
[3] A. Kurenkov et al., Appl. Phys. Lett. 110, 092410 (2017).
[4] W. A. Borders et al., Appl. Phys. Express 10, 013007 (2017).