A Novel Two-Stage Method for Deploying AI Models on RRAM-Based Compute-In-Memory System

Deploying various neural network models on CIM systems automatically is a critical part of promoting the development of CIM systems. It still remains challenges to find the optimal deployment for various algorithms on CIM systems within reasonable time. Here, we propose a novel two-stage deploying method, which combines model-guided and simulation-guided deployment methods. The method can find the optimal deployment within an acceptable optimization time. The experimental results suggested that compared to the pure analytical model-guided method, the two-stage deployment method could reduce the inference latency by 28.05%, and 72.18% for ResNet-18, and Vgg-11 respectively. Compared to the pure simulation-guided method, the optimization time could be reduced by 67.69%, and 68.84% for ResNet-18, and Vgg-11 respectively.