Impact of External Magnetic Interference on the Performance of MRAM-Based Neuromorphic Computing

The Magneto-resistive random access memories (MRAMs) have the potential to be used in neuromorphic computing based on the computing in memory (CiM) architecture. However, as a spintronic device, the resistance state will be interfered by the external magnetic field, thus causing inference degradation. Error correction code or magnetic shielding could increase its immunity, but at additional cost. Neural networks have intrinsic error tolerance, which could help relax the standard of device-level immunity, but that's seldom investigated. In this work, we comprehensively conducted an evaluation of the impact of external magnetic interference on the performance of MRAM-based neuromorphic computing, and found that the error tolerance could help protect a multilayer perceptron (MLP) neural network from showing degraded accuracy. Moreover, such tolerance differs between weights in different layers and between bits with different significance. Based on those observations, we further proposed an optimization method for immunity against field interference with reduced cost. This work helps improve the usefulness of MRAM-based neural networks in practical environments.