ReRAM-based Compute-in-Memory (CiM) architectures can accelerate AI workloads. However, device-level variations in ReRAM distort the distribution of multiply-accumulate (MAC) values, thereby degrading computation accuracy. This paper proposes a non-uniform quantization (NUQ) scheme optimized by genetic algorithms (GA) to reduce MAC readout errors while maintaining high computational efficiency. The proposed NUQ method adapts to resistance variations and signal aggregation effects. Under multi-weight and aging scenarios, it effectively mitigates signal overlap and improves readout accuracy by 60%.
Xue et al. (Thu,) studied this question.