Meta details Prometheus, a gigawatt-scale Ohio supercluster scaling to 1 million GPUs
Meta has disclosed details on Prometheus, its gigawatt-scale AI supercluster in New Albany, Ohio, which trained its Muse foundational models and is expanding toward roughly 1 million Nvidia GPUs.
Meta has detailed Prometheus, a gigawatt-scale supercluster located in New Albany, Ohio, that was used to train the company's Muse family of foundational AI models, including Muse Wonder. The company brought the facility online to support advanced model training, with reports highlighting its potential influence on global AI infrastructure benchmarks and regional economic development.[1][3][4]
Prometheus already runs hundreds of thousands of Nvidia GPUs and is scaling toward roughly 1 million processors using Nvidia GB200 and GB300 systems, with plans to eventually integrate Vera Rubin systems. Managing hardware reliability at this magnitude requires significant redundancy; Meta reported that roughly 4,900 additional GPUs are needed simply to maintain a 32,768-GPU training run through hardware failures.[2][3][5]
The supercluster's network architecture relies on millions of optical components connecting facilities spread more than 100 kilometers apart. The massive scale also places unique demands on power systems, with training runs causing electrical power swings of around 40%.[2]
Key facts
- Meta's Prometheus supercluster is a gigawatt-scale facility located in New Albany, Ohio.
- The supercluster was used to train Meta's Muse foundational models, including Muse Wonder.
- Prometheus already operates hundreds of thousands of Nvidia GPUs and is scaling toward roughly 1 million GPUs using GB200, GB300, and future Vera Rubin systems.
- About 4,900 extra GPUs are needed to support a 32,768-GPU training run through hardware failures.
- The infrastructure uses millions of optical components to link facilities located more than 100 kilometers apart.
- Prometheus training workloads cause power swings of approximately 40%.
Sources · 5 sources
- CB
Crypto BriefingArticle ·
Meta brings Prometheus supercluster online in Ohio to train its Muse AI models Meta's Prometheus supercluster enhances AI capabilities, potentially influencing global AI infrastructure standards and regional economic growth. The post Meta brings Prometheus supercluster online in Ohio to train its Muse AI models appeared first on Crypto Briefing .
Open source - SB
Shay Boloor@StockSavvyShayPost on X ·
$META just revealed that its Prometheus supercluster (which trained its new Muse AI models) already runs hundreds of thousands of $NVDA GPUs and is scaling toward roughly 1 million using GB200, GB300 and eventually Vera Rubin systems. And the scale of this infrastructure is pretty insane because Meta needs ~4,900 additional GPUs to support a 32,768-GPU training run through hardware failures. That's on top of millions of optics connecting facilities more than 100 km apart and power swings of ~40% during training showing how Meta's AI expansion is driving demand well beyond GPUs into networking, energy storage and liquid cooling.
Open source - TE
TechmemeArticle ·
Meta details Prometheus, its gigawatt-scale supercluster in Ohio, which will have ~1M GPUs at full scale and was used to train its Muse foundational models (Engineering at Meta) Engineering at Meta : Meta details Prometheus, its gigawatt-scale supercluster in Ohio, which will have ~1M GPUs at full scale and was used to train its Muse foundational models — - With the recent release of the Muse family of foundational models, we want to take you inside the infrastructure that has made it all possible.
Open source - EV
Evan@StockMKTNewzPost on X ·
Meta Platforms $META posted this today: “We trained our Muse foundational models, including Muse Wonder, at Prometheus, our gigawatt-scale supercluster in New Albany, Ohio.” https://t.co/2Cs8JwBjsg
Open source - TE
Techmeme@TechmemePost on X ·
Meta details Prometheus, its gigawatt-scale supercluster in Ohio, which will have ~1M GPUs at full scale and was used to train its Muse foundational models (Engineering at Meta) (Visit Techmeme dot com for the link and full context!)
Open source