Meta completed the first phase of Prometheus, its gigawatt-scale supercluster. It now trains frontier models like Muse Wonder. The cluster spans hundreds of thousands of GPUs across interconnected buildings. It uses two NVIDIA GPU platforms, Catalina and Clemente, with a third coming. Catalina deploys GB200 NVL72 in a two-rack, 72-GPU configuration. Clemente uses GB300 NVL72, packing 72 GPUs into a single rack. Both platforms use closed-loop liquid cooling for efficiency. Prometheus achieved over 85% goodput for pre-training and over 90% for RL runs. MTBI reached over 54K GPU hours after firmware fixes. The cluster connects GPUs via a dedicated Back-End network for direct communication. Building-level fabrics use Ethernet-based designs with up to 20,736 GPUs. The Backend Aggregation network spans 2 km to over 100 km. Meta co-engineers reliability with NVIDIA, sharing telemetry and running joint triage.
What changed
New Albany site ran web traffic before hardware, networking and building redesign for AI training.
What it unlocks
Training larger models and longer context windows than previously practical.
- ~1M GPUs at full scale
- 85%+ goodput for pre-training
- 90%+ goodput for RL runs
- ~54K GPU hours MTBI
Sources