Meta details Prometheus gigawatt supercluster: 85%+ pretraining goodput for Muse
AIatMeta · x · 2026-10-20
Meta's engineering blog details Prometheus, its gigawatt-scale supercluster in New Albany, Ohio, which trained the Muse family of foundational models including frontier model Muse Wonder.
- Phase one is complete; Meta is scaling toward 1 million GPUs, with closed-loop liquid cooling
- At hundreds of thousands of GPUs, a single underperforming host can slow an entire run. Prometheus achieved 85%+ goodput on pretraining, 90%+ on RL, and 47K GPU hours mean time between interruptions during Muse Wonder training
- The site was a conventional web-traffic datacenter two years ago, reworked for frontier training
A rare first-party look at reliability engineering at gigawatt scale.
Related event: Meta Unveils Gigawatt Prometheus Supercluster, Targets One Million GPUs(4 posts)→
More from Infra
- NVIDIA's first GB300 FP4 counter-submission still 1.64x worse perf per dollar than TPUv7 — xennygrimmato_ · 2026-10-20
- Long queues at RTX Spark SF launch for local Qwen demos built with ODS software — TheZachMueller · 2026-10-20
- Dream-1 Beta runs robot tasks at $1.72 per success, up to 41x cheaper than LLM agents — gharik · 2026-10-20
- Arm AGI CPU posts 2x end-to-end agentic workflow performance vs leading x86 in Signal65 tests — ryanshrout · 2026-10-20
- Loop engineering is a teeter-totter: you never have enough CPU cores and inference at once — kieranklaassen · 2026-10-20
- Siemens and NVIDIA push toward the first fully AI-powered factory: 85% less engineering effort — lukas_m_ziegler · 2026-10-20