Evaluation of Single Event Effects in SRAM and RRAM Based Neuromorphic Computing System for Inference

In this work, single event effects (SEEs) are analyzed in SRAM- and RRAM-based neuromorphic computing systems. SPICE simulation is employed to model single event upset (SEU) at the array level. Then SEU effects are mapped to the weight pattern change of a multi-layer perceptron (MLP), a representative artificial neural network for MNIST handwritten digit recognition. Simulations show that the RRAM-based MLP has less susceptibility to SEEs compared with the SRAM architecture. Improvements to the SRAM-based MLP reliability may be achieved by lowering bit-width and enlarging network size.