Meta details Prometheus gigawatt cluster: 100km interconnects, ~1M GPUs, 99% fewer training interruptions
ryanshrout · x · 2026-10-19
Meta has published an engineering deep-dive on Prometheus, its gigawatt-scale supercluster in Ohio that trained the Muse models and is scaling toward roughly 1 million GPUs operating as a single machine across buildings linked over 100 km.
Key details:
- Joint engineering with NVIDIA cut training interruptions 99% from the mid-2025 peak
- Each NVL72 rack runs a 64-GPU slice with eight in-rack spare GPUs
- All GPU platforms in the cluster are NVIDIA-based: GB200 and GB300 today, Vera Rubin NVL72 next
AMD executive Ryan Shrout notes Meta's AMD footprint (millions of EPYC CPUs and a 6 GW Instinct agreement for inference) is unaffected, as compute demand still outruns what any single supplier can deliver.
More from Infra
- Open-source ai-memory gives Claude Code, Codex, Cursor and 20+ AI coding tools a shared memory — alex_verem · 2026-10-19
- OpenRouter Data Shows US Model Token Use Flat for 8 Weeks as Chinese Providers Hit 79.2T Weekly — SumitGup · 2026-10-19
- CMU seminar digs into Floe lakehouse engine: no buffer pool, I/O faster than io_uring — sh_reya · 2026-10-19
- Data centers use just 0.85% of global primary energy, by the numbers — DarinFeinstein · 2026-10-19
- vLLM 0.32.0 adds structured-decision API reading first-token probabilities — AccBalanced · 2026-10-19
- GPU pricing is now set by cost of capital, not spot rates, argues Maya — AccBalanced · 2026-10-19