Meta Unveils Prometheus AI Supercluster in Ohio, Aims to Train Muse Models at Gigawatt Scale
Key Summary
Meta has launched the first phase of its Prometheus AI supercluster in Ohio, designed to train its Muse AI models at a gigawatt scale. The facility is built to run at 1 GW and will be scaled up to 1 million GPUs and 1 GW capacity. Meta reports goodput of over 85% during pre-training runs and a mean time between interruptions of 47,000 GPU hours.
Precedent & Market Context
Meta's AI efforts have been gaining momentum in recent years, with the company investing heavily in AI research and development. In 2017, Meta began its New Albany project, which has since become a hub for AI research and development. The company has also been investing in its AI infrastructure, including the development of its own AI chips.
Main Event
Meta has switched on the first phase of Prometheus, its AI supercluster at the company's data center campus in New Albany, Ohio. The facility is built to run at gigawatt scale and will be used to train Meta's Muse family of foundational models.
Analysis & Market Impact
The Prometheus supercluster is designed to improve Meta's AI performance and efficiency. The facility's use of closed-loop liquid cooling and high-capacity Ethernet connections will help to improve the reliability and efficiency of the training process. Meta's reports of goodput of over 85% during pre-training runs and a mean time between interruptions of 47,000 GPU hours indicate that the facility is performing well.
Implications & Outlook
The launch of Prometheus marks an important milestone in Meta's AI efforts. The facility's gigawatt scale will allow Meta to train its Muse models at a level that was previously not possible. The company's investment in Prometheus is also seen as a sign of its commitment to the development of its AI technology.