Periodic Labs details its stack: 4.1x Megatron throughput, frontier-beating science models on 1,300 H200s
hsu_byron · x · 2026-09-18
Periodic Labs published details of its AI infrastructure: starting from open weights, its final training run peaked at just 1,300 H200 GPUs across midtraining and RL phases, producing specialized models that Pareto-dominate frontier models like GPT-6 Astra and Claude Fable 5.1 on X-ray diffraction evaluations, now used to search for superconductors and magnets in its high-throughput labs.
Technical highlights:
- Built on Megatron, SGLang, Miles, and Ray, heavily modified for scientific RL workloads where rollouts spend hours reasoning and running tools while training steps take minutes.
- 4.1x training throughput vs Megatron on the same GPUs; SGLang contributions made inference 2.5x faster for their use case.
- Custom sandboxing, GPU memory efficiency, and capacity sharing with scientific simulation workloads sustaining 95%+ cluster utilization.
More from Infra
- Who Opens Up the Compute Middle Market? The $10-100M Deployment Standards Layer — AccBalanced · 2026-09-18
- Announced data centers hit 342GW vs 50GW installed, implying $26.3T in spend — BenBajarin · 2026-09-18
- Hyperbolic hires quant researchers to build GPU compute as a tradable asset class — YiMaTweets · 2026-09-18
- Crusoe raises $3.9B Series F at $30.9B valuation to fuel AI energy buildout — beffjezos · 2026-09-18
- Bonsai quant hits 50 tok/s at 128k context on a 24GB card, letting users run two sessions at once — julianharris · 2026-09-18
- Jeff Dean: a handful of workloads will dominate world compute, 'crying out' for specialized silicon — AccBalanced · 2026-09-18