Periodic Labs: 1,300 H200s beat frontier models on X-ray evals, 4.1x Megatron throughput
zephyr_z9 · x · 2026-09-16
Periodic Labs details its AI infrastructure: starting from open weights, a peak of just 1,300 H200 GPUs across midtraining and RL produced specialized models that Pareto-dominate frontier models like GPT-6 Astra and Claude Fable 5.1 on X-ray diffraction evals. Their Megatron-derived stack delivers 4.1x training throughput, SGLang contributions made inference 2.5x faster, and cluster sharing with simulation workloads sustains 95%+ utilization — all aimed at discovering better superconductors and magnets.
More from Infra
- Brad Gerstner: AI buildout hinges on revenue keeping its steep growth curve — markjeffrey · 2026-09-16
- Reported: NVIDIA hardware delivers 40% more throughput at the same power — karlfreund · 2026-09-16
- NVIDIA Shares New Vera Platform Data at AI Infra Summit — karlfreund · 2026-09-16
- Token usage forecast to hit 4 quintillion by 2030, agentic AI to drive most of it: Evercore — Beth_Kindig · 2026-09-16
- NVIDIA's Ian Buck shares Groq inference performance data at AIInfraSummit — karlfreund · 2026-09-16
- MLPerf Inference v6.1 results imminent: 30 submitters, new accelerators and platforms — TheKanter · 2026-09-16