MSCA: Model-Driven Search for Optimal Configuration for SpMM Accelerators

Sparse Matrix-Matrix multiplication (SpMM) is a cornerstone operation across AI algorithms, notably in advanced machine learning models such as transformers and graph neural networks. While various SpMM accelerators have been developed, optimal system configuration search method remains an underexplored area. Traditional exhaustive search methods are time-consuming and lack deep system understanding. This paper introduces a model-driven approach to efficiently search for optimal configurations. We transform the configuration search into a mixed integer problem, where we simplify sparsity impact with random experiments and a lookup table. For the first time, we model latency considering both simple and double buffering. Our approach allows flexible design goal setting for different scenarios. Experimental results show our model-derived configuration can outperform almost all counterparts in latency and it can effectively indicate the system's minimum area requirements.