Framework and Quantization Method Design for CIM

This paper proposes a framework for simulating RRAM-based CIM. The framework is designed to explore changes in accuracy under the constraints of ADC precision, resistance on/off ratio, and activated WL number. We also propose an accuracy-aware quantization methodology to reduce the number of quantization bits required. Experiment results show that the quantized IFMs only reduce the accuracy by 0.09%, while the accuracy of quantized weights dropped by 10.55%. In the future, we will add R on/off ratio and current error to further study the relationship between number of activated WL and CIM accuracy.