Design of spike-timing-dependent plasticity synapses based on CoPt-SOT device and its application in all-spin spiking neural network

Abstract

Spintronic could be used to simulate synapses or neurons due to its multistate storage characteristics. In this work, a reliable design of all-spin spiking neural networks (SNN) based on spin–orbit torque (SOT) devices has been proposed in A1 CoPt single layer. The CoPt-SOT devices exhibited field-free SOT switching, and the magnetization reversal mechanism was inferred to be a combination of domain nucleation and domain-wall propagation as observed through magneto-optical Kerr microscopy images. Moreover, the current-induced SOT switching process of the device exhibited stable multistate magnetic switching behavior, which can be controlled by varying the amplitude and pulse width of the current pulse. Meanwhile, the spike-timing-dependent plasticity (STDP) curve was inverted when the SOT switching polarity was reversed by different magnetic fields, and the change in anomalous Hall resistances (ΔRH) in the STDP curve was linearly related to the SOT switching ratio. In addition, at the zero magnetic field, we constructed an all-spin SNN using STDP synapses and leaky integrate-and-fire neurons of CoPt-SOT devices. The handwritten digits recognition rate of this all-spin SNN network was 89.9%. These results substantiate that the CoPt single layer represents a promising hardware solution for high-performance neuromorphic computing, with applicability in the domain of SNN.