Effect of process variability in spin orbit torque magnetic tunnel junctions on the performance of spiking neural networks

Abstract

Emerging non-volatile memory devices such as spin-orbit torque based magnetic tunnel junctions (SOT-MTJ) have shown promise has suitable candidates for universal memory due to their low power consumption and high speed opera- tion. However, the properties of SOT-MTJ devices are not limited to memory applications only, rather they have also been explored as neurons and synapses in neuromorphic computing. Particularly, brain inspired spiking neural network models have been developed that are supported by SOT-MTJ based hardware implementations. Such systems show capabilities of in-memory computing thus preventing the memory wall problem observed in conventional Von Neumann architecture. However, the state-of-the-art implementations of spiking networks based on SOT-MTJ devices do not consider device-to-device variations and all the devices used in a network are assumed to be completely identical to one another. To solve cognitive tasks, the algorithms of these networks are bound to become more complex and their size will need to scale up in order. As a result, the number of devices will also rise up exponentially making device level variations a very practical consideration. In this study, we have modeled spik- ing neural networks based on SOT-MTJ based neurons and synapses. We have benchmarked the performance of the network on the MNIST handwritten digit recognition dataset. Firstly, we have trained our network offline and tested it on the MNIST testset considering device level variations of the SOT-MTJ based neuron. We have shown that as the standard deviation of geometric and material parameters of the SOT-MTJ increase to 10%, the accuracy of the SNN can drop by as much as 27%. We have also observed that the performance of the network is more susceptible to the variations in the neurons in the output layer compared to the hidden layer. As the parameters of the output layer neurons change, it makes the system imbalanced in predicting one particular digit. The network has also been found to show varying degree of performance for same degree of variation in different parameters showing drop in performance at a greater speed for free layer thickness compared to damping. We have also trained the net- work online considering a two layer spiking neural network using spike timing dependent plasticity. In this case, we have varied the geometric and material pa- rameters of the domain wall based magnetic tunnel junction used as the synapse. We have shown that as the parameters of the synapse changes, it either slows down or speeds up its learning capabilities. If the saturation magnetization is increased, it becomes more difficult to change the domain wall position, so it slows down the weight update and eventually the network ends up not learning the patterns well enough, hence the accuracy also degrades in this case. Thus, we have considered the variability of SOT-MTJ based neurons and synapses in spiking neural networks and analyzed the impact of device level variations on system level performance.