Emerging Trends and Advances in Computing-in-Memory: A Software-Hardware Co-Design Perspective

Computing-in-memory (CIM) has emerged as a promising solution to the memory bottleneck in deep neural network (DNN) acceleration, yet maximizing its potential requires careful software-hardware co-design. This paper systematically examines key co-optimization techniques for CIM architectures, including quantization, sparsity, compact models, and decomposition. The analysis also explores how software stacks can effectively bridge the gap between high-level DNN frameworks and CIM hardware primitives. Finally, this work identifies unresolved challenges across the hardware-software stack and propose concrete research directions to advance this emerging paradigm.