MLPerf v6.1: GPT-OSS-20B trains in 19.7 minutes on GB300 NVL72, 2.86x faster than H100
TheZachMueller · x · 2026-11-17
Lambda reported MLPerf Training v6.1 results: on a single NVIDIA GB300 NVL72 system (64 Blackwell Ultra GPUs), GPT-OSS-20B trained in just 19.697 minutes — 2.86x faster than 64 H100s (Cisco) last round and 1.37x faster than 64 Blackwell GPUs (CoreWeave).
- First multimodal benchmark run: in an unofficial reproduction, Flux.1 trained in 26.36 minutes on the same system, 1.50x faster than last round's fastest on that hardware
- Takeaway: new hardware shortens training, while better software keeps finding speed on hardware you already have
More from Infra
- Anthropic's annualized revenue hits ~$81B ahead of IPO, with $895B compute commitments — lennysan · 2026-11-17
- Sandbox startup speed is moot if compute sits 200ms from your infra — DanielLockyer · 2026-11-17
- PyTorch Foundation Honors 15 Winners in Fourth Annual Contributor Awards — PyTorch · 2026-11-17
- NVIDIA ops chief Deb Shoquist wins GSA's 2026 Woman of Innovation Award — nvidia · 2026-11-17
- Sam Altman walks back RAM comments: 'I want everyone to have a lot of RAM' — The Verge AI · 2026-11-17
- Triangle Launches Cross-Model KV Cache Transfer, Up to 10.8x Faster TTFT at 1M Context — ycombinator · 2026-11-17