Impact of Process Variation in Spin–Orbit Torque-Based Magnetic Tunnel Junctions on the Performance of Spiking Neural Networks

Spiking neural network (SNN) with neurons and synapses made of spin–orbit torque (SOT) and domain wall motion-based magnetic tunnel junctions (MTJs) offer a low-power alternative to deep neural networks (DNNs). The present study demonstrates the effect of device-level variations in such neurons and synapses on the performance of a three-layer SNN trained offline using supervised learning and a two-layer SNN trained online using unsupervised spike-timing-dependent plasticity (STDP). The spintronic properties of the neuron and the synapse have been modeled based on the stochastic Landau–Lifshitz–Gilbert (LLG) equation, using macro-and micro-magnetic simulation techniques, respectively. The variations in thickness, saturation magnetization, and damping of the free layer (FL) along with spin Hall angle of the metal have been considered in variability analysis. The results of our analysis show that in both offline and online trained systems, the accuracy of the SNN drops by as much as 20%, for 10% standard deviation of device parameters. The offline trained network is observed to be more sensitive to the variations in neurons located in the output layer. In the online trained network, performance degradation can be traced back to the correlation between the distribution of learned weights and synaptic parameter variation. Among the device parameters considered, the variation in spin Hall angle of the metal layer is observed to have the most significant impact on the performances of the SNNs.