Skip to content
Neural Daily

BusinessNews

Meta Prometheus aims for ~1M GPUs in Ohio

On October 19 Meta published a detailed Engineering post on Prometheus, its gigawatt-scale supercluster in New Albany, Ohio. Phase one is complete and the target is about one million GPUs. Muse Wonder models trained there. Pre-training goodput exceeded 85% and RL goodput exceeded 90%.

By: Mikola Máté Patrik · · 2 min read

Rows of server racks with yellow cabling along an aisle of Meta's Prometheus supercluster in New Albany, Ohio
Image: Meta · Press kit · Official Meta Engineering blog image · Source

What gigawatt-scale means here

Prometheus already runs hundreds of thousands of GPUs across interconnected buildings as one machine. Full scale targets roughly 1 million distributed GPUs. Two years ago New Albany was a conventional web traffic site. It has been rebuilt for frontier training with closed-loop liquid cooling (Meta Engineering).

Hardware generations

Two NVIDIA platforms run today, with Vera Rubin NVL72 arriving soon. Catalina uses the GB200 NVL72 Ariel variant: two racks, 36 GPUs per rack linked by NVLink into a 72-GPU domain, with 800G scale-out per GPU (2× ConnectX-7). Clemente moves to GB300 NVL72: 288 GB HBM per GPU (1.5× GB200), native 800G ConnectX-8, and the full 72-GPU domain in a single rack.

Each building fabric interconnects up to 20,736 GPUs on an Ethernet back-end. The Backend Aggregation (BAG) network links building fabrics over 2 km to more than 100 km using Arista 7800 chassis.

Meta chart of GB200 and GB300 interruptions on-prem and in the cloud from June 2025 to September 2026, showing a steep decline
Meta · Press kit · Official chart from the Meta Engineering blog post · Source

Source: Meta. GB200 and GB300 interruptions falling from June 2025 to September 2026, on-prem and in the cloud.

Reliability: a 99% drop in interruptions

In mid-2025 interruptions peaked near 1,000 events per 8K GPUs per day. After a year of work Meta reports a 99% reduction. Muse Wonder pre-training goodput exceeded 85% and RL goodput exceeded 90%, with wall-clock MTBI around 47k GPU hours. Firmware fixes later pushed MTBI above 54k GPU hours.

NVL72 expands the fault blast radius 9× versus NVL8. Each domain therefore runs a 64-GPU slice and keeps eight in-rack spares. A 32,768 active-GPU job provisions 523 NVL72 racks, or 37,656 GPUs, about 13% above the active footprint. PAFT isolates a faulty data-parallel replica while the rest keep training. Larger faults restart from checkpoints.

Power and what comes next

Synchronized phases across tens of thousands of GPUs create 30-40% power swings. Dynamo Dimmer gradually caps GPU power near the electrical ceiling. The San Miguel Vera Rubin NVL72 platform is next. Hyperion, the upcoming Louisiana supercluster, is being designed from Prometheus lessons.

    Produced with the help of AI, reviewed by an editor. How we make our articles

    Reviewed and edited by the Neural Daily desk. About

    Related articles