Matmul Optimization: Using Data Movement in Registers as Cost Function

cosminnegruseri · x · 2026-08-19

Asserts that matmul performance is dominated by data movement cost, not multiplication count. Shares a benchmark for 4x4 matmul optimization by @yaroslavvb, where the cost function is driven by manhattan distances for moving data in registers. Considers this closer to optimization with practical impact. Includes a GitHub project link.

Related event: Matmul optimization is bounded by data movement, not multiplication count(4 posts)→

Original post →

More from Infra

Infra channel →