New B300 FP4 Attention Kernels Deliver 1.69x Speedup
tuananh_org · reddit · 2026-07-14
Someone shared a set of new FP4 attention kernels for the B300, claiming up to a 1.69x speed improvement compared to FA4.
This update is fundamentally about inference/operator-level performance optimization rather than the models themselves. If accurate, it indicates substantial room for optimizing attention implementations on new hardware platforms, particularly along low-precision pathways.
More from Infra
- AWS on agent prompt caching: explicit + implicit patterns to slash token costs — AWS ML Blog · 2026-10-27
- Thieves spoofing GPS steal $100M+ in data center cargo since June — Ars Technica AI · 2026-10-27
- Lam Research: weekly token usage up ~250x in 18 months — Beth_Kindig · 2026-10-27
- AWS shows how to build enterprise agentic retrieval on Bedrock with CDK — AWS ML Blog · 2026-10-27
- Encrypted GPT-2 speedrun challenges FHE inference: best is 3.031 s/token, goal sub-1s — tensorqt · 2026-10-27
- Open-source Latch compiles PyTorch into training megakernels, 2.75x faster on H100 — sguada · 2026-10-27