Matmul Optimization Bottleneck: Data Movement, Not Multiplications

yaroslavvb · x · 2026-08-19

Discusses the bottleneck in matmul performance optimization. Notes that the PRAM model is obsolete due to wire delays. Modeling these delays (as in arxiv 2205.04934) shows that the matmul lower bound is known and achieved by systolic arrays. Performance in modern computers is dominated by data movement costs, not the number of multiplications.

Related event: Matmul Optimization Is Bound by Data Movement, Not Multiplication Count(2 posts)→

Original post →

More from Infra

Infra channel →