Characterization and Mitigation of Relaxation Effects on Multi-level RRAM based In-Memory Computing

In this paper, we investigate the relaxation effects on multi-level resistive random access memory (RRAM) based in-memory computing (IMC) for deep neural network (DNN) inference. We characterized 2-bit-per-cell RRAM IMC prototypes and measured the relaxation effects over 100 hours on multiple 8 kb test chips, where the relaxation is found to be most severe in the two intermediate states. We incorporated the experimental data into SPICE simulation and software DNN inference, showing DNN accuracy for CIFAR-10 dataset could degrade from 87.35% to 11.58% after 144 hours. To recover the largely degraded accuracy, mitigation schemes are proposed: 1) at the circuit level, the reference voltage for RRAM IMC could be calibrated after 80 hours when the relaxation is saturated. 2) At the algorithm level, the weights are trained with lower percentages to be quantized to the two intermediate states. With both schemes applied, the accuracy could be recovered to 87.32% for long-term stability.