Inside Prometheus: Meta’s Gigawatt-Scale Supercluster
Summary
Meta’s Prometheus is described as a gigawatt-scale AI training supercluster based in New Albany, Ohio, designed to scale toward around 1 million GPUs. The article covers hardware generations (Catalina GB200 NVL72 and Clemente GB300 NVL72), network fabrics and aggregation (BAG), reliability improvements, fault recovery, power management, cooling innovations, and plans to add Vera Rubin NVL72 systems (San Miguel) to further scale performance and efficiency.