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)→
More from Infra
- OpenAI uses ~20% compute for inference during training — eliebakouch · 2026-08-19
- Vercel KMS Lets You Sign JWTs Without Managing Private Keys — cramforce · 2026-08-19
- Ling-3.0-tiny Runs 128K Context on $249 8GB Orin Nano — Puzzleheaded_Base302 · 2026-08-19
- Apple's Foundation Model Framework: Hybrid AI Routing with Dynamic Profiles — Scobleizer · 2026-08-19
- Docling Graph turns documents into queryable knowledge graphs using Pydantic — techNmak · 2026-08-19
- CoreWeave hits $2.6B quarterly revenue in just 25 quarters, a milestone AWS took 40 to reach — FinanceYF5 · 2026-08-19