Meta completes first phase of its gigawatt-scale Prometheus supercluster

What was announced

Meta said it has completed the first phase of its gigawatt-scale Prometheus supercluster at its New Albany, Ohio data center. According to the company, this capacity was used to train its new Muse foundation models.

Prometheus currently spans hundreds of thousands of GPUs across multiple buildings. The site is designed to function as a single machine, linked by miles of fiber. Meta aims to reach roughly 1 million GPUs at full scale.

Hardware and cooling

The supercluster currently uses two NVIDIA GPU platforms, with a third based on NVIDIA's Vera Rubin NVL72 arriving soon. The first platform, Catalina, uses the Ariel variant of GB200 NVL72, a two-rack configuration that links 72 GPUs into a single accelerator domain.

Meta also points to technologies such as closed-loop liquid cooling for efficiency, along with its Open Rack V3 open hardware design.

Training efficiency

By focusing on system stability, Meta says it achieved more than 85% goodput for pre-training and more than 90% for reinforcement learning runs of Muse Wonder. Mean time between interruptions rose from about 47k GPU hours to more than 54k GPU hours after firmware fixes.