Periodic Labs beats frontier models on materials analysis with just 1,300 H200s
LiamFedus · x · 2026-09-16
Periodic Labs (with ex-OpenAI's Liam Fedus) detailed its AI infrastructure: starting from open weights, a mid-training plus RL run peaking at only 1,300 H200 GPUs produced a model called Neon that Pareto-dominates GPT-6 Astra and Claude Fable 5.1 on its X-ray diffraction benchmark. The stack builds on Megatron, SGLang, Miles and Ray—delivering 4.1x Megatron training throughput and 2.5x faster inference—plus custom sandboxing and 95%+ cluster utilization shared with scientific simulation workloads. High-throughput labs in Menlo Park close the loop: fresh experimental data trains models that pick the next experiments, targeting superconductors, magnets and semiconductors.
More from Infra
- NVIDIA ships NVFP4-quantized GLM-5.3-Flash, trending on Hugging Face — nvidia · 2026-09-16
- "Put a datacenter on top of the Salesforce tower": a joke born of AI's land rush — matt_slotnick · 2026-09-16
- Ben Bajarin 'supremely bullish' on AI buildout after Pat Gelsinger podcast — BenBajarin · 2026-09-16
- Benzi coding agent hits 78.2% SWE-bench reading far less code than Claude Code — DonkeyTheKing · 2026-09-16
- Vera Rubin NVL72 debuts in MLPerf v6.1 with up to 3.7x throughput vs GB300 NVL72 — nordicinst · 2026-09-16
- MLPerf v6.1 debuts agentic benchmark and first 1T+ parameter model run — TheZachMueller · 2026-09-16