Spintronic neuromorphic hardware using domain wall-based neurons and quantized synapses

Abstract We simulate the functionality of an artificial neuron and synapse using spin-orbit torque-based spintronic devices and implement a fully connected artificial neural network (ANN). These neuro-synaptic devices are emulated via transverse domain wall dynamics within a magnetic nanotrack comprised of heavy-metal (HM)/ferromagnet (FM) heterostructures, parameterized using Pt-based alloys, and CoFeB for the HM layer and FM layer, respectively. The ReLU activation function of the neuron is mimicked using domain wall (DW) motion. The synapse is modeled using current-induced DW dynamics with a corrugated HM/FM nanotrack, where the electrical conductance of the pinned DW across pinning points acts as stable synaptic weights. The fully connected ANN, using conventional float32 weights, achieves an accuracy of ~97% and ~86% for the MNIST and Fashion-MNIST datasets, serving as a test bed for our simulations. To implement a sparse, low-memory-footprint ANN, we quantize the trained weights into discrete levels, achieving an accuracy of ~95% and ~62% for MNIST and Fashion-MNIST, respectively, which is restored back to the baseline accuracy after fine-tuning. We demonstrate the write and read energies of the ANN are on the order of pJ, with an operational latency of 8 ns.