Inside Meta Prometheus: The Gigawatt-Scale Supercluster Scaling to 1M GPUs

Engineering at Meta has publicly disclosed the comprehensive architectural specifications for 'Prometheus', its flagship gigawatt-scale artificial intelligence supercluster situated in New Albany, Ohio. The multi-campus installation represents the primary training facility behind Meta's next-generation open frontier models, including the Muse architecture family and the frontier reasoning engine Muse Wonder.

The New Albany site—which operated two years ago as a standard web tier data center—has completed its first operational deployment phase. Prometheus currently runs hundreds of thousands of synchronized GPU accelerators distributed across adjacent purpose-built halls and rapid-deployment facilities, engineered along an infrastructure runway that expands toward 1 million active GPUs operating as a unified compute fabric.

Tri-Generation Hardware Platforms: Catalina, Clemente, and San Miguel

The physical compute footprint of Prometheus is built entirely upon dense, liquid-cooled NVIDIA server infrastructure across three progressive platform tiers:

51.2T RoCE Network Fabric & Backend Aggregation (BAG)

To interconnect hundreds of thousands of GPUs without standard vendor lock-in, Meta standardized on an open RoCEv2 (RDMA over Converged Ethernet) fabric rather than proprietary InfiniBand. Within each hall, non-blocking leaf-spine fabrics cluster up to 20,736 GPUs utilizing state-of-the-art 51.2 Tbps merchant silicon switches.

To bridge workloads across multiple physical campus structures separated by distances ranging from 2 kilometers to over 100 kilometers, Meta engineered an inter-facility aggregation layer termed Backend Aggregation (BAG), preserving near-zero latency penalty during distributed all-to-all collective communications.

Production Reliability: 54,000 GPU-Hour MTBI and >85% Training Goodput

At cluster scales exceeding hundreds of thousands of accelerators, operational success is governed entirely by system reliability. Meta reported that training 'Goodput'—the ratio of active training time generating valid gradient updates without being wiped out by infrastructure faults—surpasses 85% for foundational pre-training and exceeds 90% during reinforcement learning runs on Muse Wonder.

Mean Time Between Interruptions (MTBI) increased from 47,000 GPU-hours to more than 54,000 GPU-hours following collaborative firmware optimizations with on-site NVIDIA resident engineers. Fault frequency plummeted 99% from mid-2025 peak failure rates of nearly 1,000 incidents per 8,000 GPUs per day down to a target of just 2 incidents daily.

Managing 30–40% Instantaneous Grid Power Swings with 'Dynamo Dimmer'

Synchronous training on massive accelerator arrays induces severe physical power fluctuations: collective barriers and computational handoffs trigger sudden 30% to 40% drops and spikes in electrical demand (tens of megawatts within milliseconds), threatening local utility transmission grids.

To prevent catastrophic grid trips, Meta engineers deployed proprietary 'Dynamo Dimmer' orchestration software. The system proactively throttles clock frequencies and smooths ramp rates, working in tandem with high-rate capacitor banks, in-rack battery backup units (BBUs), and automated telemetry integration with municipal power operators.

“At gigawatt scale, the laws of data center physics take precedence over algorithmic design. Frontier model dominance is no longer won merely by buying compute, but by mastering power-ramp dynamics, firmware resiliency, and fault isolation under continuous load.”

The operational blueprint of Prometheus establishes Meta not only as an unprecedented buyer of silicon, but as a defining benchmark for high-reliability hyperscale infrastructure in the multi-gigawatt era.