Meta details Prometheus, its gigawatt-scale AI supercluster with ~1 million GPU target

Meta published technical details of Prometheus, its largest AI infrastructure project to date, located in New Albany, Ohio. The first phase is complete and was used to train the Muse and Muse Wonder models. The cluster currently houses hundreds of thousands of GPUs across multiple buildings and is designed to reach approximately 1 million GPUs at full scale. Meta achieved over 85% goodput for pre-training and over 90% for reinforcement learning runs, with mean time between interruptions improving from ~47K to over 54K GPU hours through firmware fixes.

Hardware generations and network architecture

• Catalina (GB200 NVL72): Two-rack Ariel variant with 72 Blackwell GPUs, 13.4 TB HBM3e, 720 peak PFLOPS FP8, 800G I/O per GPU via dual ConnectX-7 400G NICs, ~120 kW total power. Uses Meta's Open Rack V3 (ORv3) high-power rack design contributed to OCP.
• Clemente (GB300 NVL72): Single-rack 72 Blackwell Ultra GPU platform with 20.7 TB HBM3e (288 GB/GPU), over 1 exaFLOP FP4, native 800G per GPU via ConnectX-8, 2:1 GPU-to-CPU ratio, ~135 kW rack power, 5-10% performance improvement over GB200.
• Third platform: NVIDIA Vera Rubin NVL72 arriving soon.
• Network: Dedicated Ethernet-based Back-End (BE) network for GPU-to-GPU communication, separate from standard data center network. Building-level fabrics interconnect up to 20,736 GPUs; low-latency fabrics within buildings, aggregation network between buildings.
• Cooling: Closed-loop liquid cooling; Clemente supports modular Air-Assisted Liquid Cooling (AALC) or Facility Liquid Cooling (FLC).

Why it matters

Meta's contribution of its Open Rack V3 design to the Open Compute Project provides a reference for high-power AI rack infrastructure, though the GPUs and much of the networking remain proprietary NVIDIA technology.