Voltage Gated Domain Wall Magnetic Tunnel Junction-based Spiking Convolutional Neural Network

TL;DR

This paper introduces a voltage-gated domain wall magnetic tunnel junction device driven by spin-orbit torque, enabling efficient neuromorphic computing with high linearity and over 85% accuracy on CIFAR-10 pattern recognition.

Contribution

It presents a novel voltage-gated DWM-based MTJ device that eliminates the access transistor and demonstrates its application in spiking neural networks.

Findings

Device achieves highly linear synaptic behavior.

Linearity depends on DMI and temperature.

Achieves over 85% accuracy on CIFAR-10.

Abstract

We propose a novel spin-orbit torque (SOT) driven and voltage-gated domain wall motion (DWM)-based MTJ device and its application in neuromorphic computing. We show that by utilizing the voltage-controlled gating effect on the DWM, the access transistor can be eliminated. The device provides more control over individual synapse writing and shows highly linear synaptic behavior. The linearity dependence on material parameters such as DMI and temperature is evaluated for real-environment performance analysis. Furthermore, using skyrmion-based leaky integrate and fire neuron model, we implement the spiking convolutional neural network for pattern recognition applications on the CIFAR-10 data set. The accuracy of the device is above 85%, proving its applicability in SNN.