Challenges with Hardware-Software Co-design for Sparse Machine Learning on Dataflow Architectures

Explore the challenges of hardware-software co-design for sparse machine learning on dataflow architectures in this 10-minute conference talk from ACM SIGPLAN. Delve into the problem landscape arising from using general tensor algebra accelerator frameworks for real-world machine learning applications. Discover three key challenges for correctness and performance: supporting tensor reshaping and nonlinear operations, optimizing dataflow through kernel fusion and optimal ordering, and leveraging sparsity structure. Understand the motivation behind addressing these issues in domain-specific languages, compiler frameworks, and architectural designs for sparse machine learning. Learn how researchers extended the Sparse Abstract Machine, a general tensor algebra compiler and architectural model, to identify these crucial challenges in real-world sparse machine learning models.