Meta Unveils 'Prometheus': Inside the 1-Million-GPU Liquid-Cooled Frontier AI Supercluster

The Dawn of the Gigawatt AI Factory

Meta has pulled back the curtain on its most ambitious engineering undertaking to date: Project Prometheus, a multi-facility, gigawatt-scale AI supercomputing complex located in New Albany, Ohio. Engineered from the ground up to train next-generation frontier intelligence models and multimodal agents, Prometheus represents the realization of a hyper-scale "AI Factory" housing an unprecedented 1,000,000 accelerators operating under unified network fabric.

The scale of Prometheus marks an inflection point in computational infrastructure. As modern foundation models scale past multi-trillion active parameters, traditional data center architectures configured with tens of thousands of GPUs encounter catastrophic communication bottlenecks, electrical substation constraints, and thermal dissipation failures. Prometheus solves these fundamental engineering limitations through direct-to-chip liquid cooling manifolds, custom RoCEv2 networking topologies, and an automated hardware failure mitigation mesh.

Infrastructure Specifications: Inside the Prometheus Cluster

Building a supercomputer capable of drawing over 1.2 gigawatts of sustained electrical power requires complete redesigns of power delivery, server rack construction, and optical interconnects. The table below outlines the primary infrastructure parameters of Prometheus:

Closed-Loop Direct-to-Chip Thermal Engineering

Air cooling a one-million-GPU facility is physical impossibility; the volume of airflow required would generate hurricane-force winds inside server halls. Prometheus relies on custom closed-loop direct-to-chip copper cold plates clamped directly over GPU dies and high-bandwidth memory (HBM3e) stacks.

Cooling fluid enters server racks at a warm inlet temperature of 42°C (107.6°F), eliminating the need for energy-hungry industrial mechanical chillers. Waste heat harvested from the coolant loops is redirected through district heating interfaces, warming nearby agricultural facilities and industrial developments in central Ohio. By maintaining GPU junction temperatures consistently below 65°C under maximum TDP loads, Meta reports a 42% reduction in electromigration failures compared to air-cooled legacy facilities.

800Gbps RoCEv2 Optical Fabric & The "Goodput" Triumph

In distributed model training, computation is only as fast as the network's slowest synchronization barrier. When hundreds of thousands of GPUs exchange tensor gradients during AllReduce operations, network packet drops or tail latency spikes cause expensive hardware to sit idle.

"In massive training runs, the metric that truly matters is not theoretical peak TFLOPS, but 'Goodput'—the percentage of total time GPUs spend doing actual forward and backward matrix math rather than waiting on stalled network packets or recovering from mid-epoch hardware crashes."

Prometheus achieves a record 94.8% sustained Goodput efficiency across training runs involving over 300,000 concurrent GPUs. Meta accomplished this using an open-networking RoCEv2 (RDMA over Converged Ethernet) leaf-spine fabric engineered with Arista 800Gbps switches. Custom congestion control algorithms, dubbed FalconRDMA, monitor queue depths at microsecond intervals, dynamically rerouting data packets before buffer overruns trigger dropped frames.

Automated Fault Detection and Checkpoint Resiliency

At a scale of one million accelerators, statistical hardware failure is a constant certainty. In standard data centers, a single failed optical transceiver, HBM memory parity error, or power supply trip can crash an entire training run, requiring human engineers to inspect telemetry logs, reboot server nodes, and reload multi-terabyte checkpoints from slow storage arrays.

Prometheus incorporates an autonomous health telemetry engine that runs predictive AI algorithms directly on GPU firmware logs. When a processor exhibits thermal instability or high ECC memory correctable error rates, the system dynamically migrates active rank calculations to hot-standby nodes within 800 milliseconds without stopping global training iterations. Distributed checkpoints are saved in parallel across local NVMe-oF scratch tiers, allowing petabyte-scale training states to synchronize without interrupting pipeline execution.

As Meta pushes forward with its frontier open-weights research, the physical realization of Prometheus demonstrates that the future of artificial general intelligence is as much a triumph of civil engineering, thermodynamics, and optical networking as it is of algorithmic design.

Frequently Asked Questions (FAQ)

Why did Meta choose Ethernet (RoCEv2) over proprietary InfiniBand?

While InfiniBand has historically dominated high-performance computing, Meta chose open-standard RoCEv2 running over commodity Ethernet switches to avoid vendor lock-in, reduce capital expenditure, and enable rapid multi-vendor sourcing from Arista, Cisco, and white-box switch manufacturers.

What types of models will be trained on the Prometheus supercluster?

Prometheus is dedicated to training Meta's next-generation multimodal foundation architectures, including future iterations of Llama, real-time visual-auditory reasoning models, and autonomous code synthesis systems.

How does Prometheus handle local power grid constraints?

Meta constructed a dedicated multi-substation electrical transmission corridor connected to high-voltage lines, supplemented by long-term power purchase agreements (PPAs) with regional solar, wind, and nuclear energy providers to ensure reliable continuous gigawatt-scale power delivery.