Challenges and Trends in Sparse Matrix Multiplication on HPC Workloads

This conference talk by Hector Arroyo, Research Engineer at OpenChip, explores the critical role of sparse matrix multiplication (SpMM) in high-performance computing and its implementation in RISC-V and Open Compute architectures. Discover how SpMM enables scientific simulations, machine learning, and graph analytics while potentially reducing memory usage, computation time, and energy consumption. Learn about the optimization challenges posed by matrices with predominantly zero elements, especially in large AI models and energy-efficient systems. The presentation examines hardware advancements including specialized accelerators and dynamic zero-skipping techniques, alongside software innovations such as various sparse matrix formats (CSR, CSC, ELLPACK, SellC-Sigma) and AI/ML-based optimization approaches. Explore current research addressing workload imbalance, memory access pattern optimization, and the balance between flexibility and performance. Gain insights into emerging trends and future directions for enhancing SpMM efficiency in next-generation HPC systems.