A CIM Crossbar Array Data Mapping Methodology for Unstructured Sparse CNN

CIM (Compute-In-Memory) is an effective technique to enable in-array computation and reduce data movement in modern machine learning systems. However, the irregularly distributed zeros in pruned CNN networks would cause inefficient usage of memory arrays in CIM. In this paper, we propose a CIM data mapping methodology for unstructured sparse CNN. This method flexibly rearranges non-zero weights in unstructured sparse CNN models by strategically swapping rows within the weight map. We also introduce a CIM architecture with switchable CIM macro to efficiently handle sparse weight maps. The experimental results have shown that our design attains up to 1.97x, 1.8x better performance, and 7.45x, 5.02x energy efficiency improvement than the conventional CIM solutions for unstructured sparse VGG16 and ResNet50 respectively.