Inside Meta Prometheus: The Gigawatt-Scale Supercluster Scaling to 1M GPUs
ai••2 min•engineering.fb.com
Engineering at Meta has published the technical blueprint for its gigawatt-scale Prometheus supercluster in New Albany, Ohio, engineered to interconnect up to 1 million NVIDIA GB200/GB300 GPUs with 51.2T RoCE fabrics, >85% training goodput, and Dynamo Dimmer grid stabilization.
Read full article→#Meta AI