Reza Zadeh claims faster matrix multiplication algorithm, suspects labs near exponent 2
Reza_Zadeh · x · 2026-10-07
Reza Zadeh says he has found a faster matrix multiplication algorithm, adding that with more compute he could do better still — and suspects AI labs already have something close to exponent 2 for matrix multiplication.
Matrix multiplication underpins all neural network training and inference, so any exponent improvement directly affects compute costs. Paper and verification to watch.
Related event: Reza Zadeh Claims Faster Matrix Multiplication Algorithm(2 posts)→
More from Infra
- ANVIL III Optimizer Claims 62% Pretraining Cost Cut at Frontier Scale, Beats Tuned Muon — kellerjordan0 · 2026-10-07
- H100 Supply Crunch: 4-Year-Old Servers Are Scarce, Not Cheap, Says Compute Exchange — ns123abc · 2026-10-07
- Omnigent's four-point playbook tames runaway LLM bills with smart model routing — matei_zaharia · 2026-10-07
- brrr: Open-Source GPU Stress Test for BF16, FP8, MXFP4 and NVFP4 — ypatil125 · 2026-10-07
- Kalshi bets on 'Silicon Spread': trading power-in, compute-out as AI's key commodity spread — nima_owji · 2026-10-07
- "Just give me 1.5MW": the power bill for running 10,000 Opus-quality models at home — generativist · 2026-10-07