François Fleuret on why matrix multiplication keeps GPUs perfectly busy
francoisfleuret · x · 2026-10-13
Former Meta researcher François Fleuret explains why matrix multiplication is ideal for AI compute: memory I/O is slow and float ops are fast, but matmul transfers only O(N²) values while performing O(N³) operations — so at some N, both memory bandwidth and floating point units are perfectly saturated. It's the core reason GPUs train models so efficiently.
More from Infra
- LangChain built SmithDB, a new database just for agent traces in LangSmith — hwchase17 · 2026-10-13
- IMF forecast assumes no AI productivity boost, sees little wage or job impact yet — rohanpaul_ai · 2026-10-13
- IMF Forecast: AI Adds No Measurable Productivity Yet, Bust Would Cost Twice the Boom's Gain — rohanpaul_ai · 2026-10-13
- NVIDIA ships a NURBS B-rep kernel on USD, but devs slam its nondeterministic geometry — rms80 · 2026-10-13
- Onchain Inference Grows 5x in 3 Months: 15 Teams to Watch This Cycle — 0xJeff · 2026-10-13
- Running llama.cpp on NVIDIA + AMD GPUs: build separate CUDA and HIP binaries — espece-de-bon · 2026-10-13