A Look Inside Prometheus: Designing Meta's Large AI Cluster Deployments

Author: Maggie Han, Technical Program Manager, Infrastructure, Meta & Abhinav Triguna, Technical Program Manager, Network, Meta Collection: 2025 OCP Global Summit - Expo Hall Sessions

Summary

Meta's Prometheus is a gigawatt-scale AI accelerator cluster that integrates GPU platforms, storage, compute, and networking into a unified open infrastructure. The presentation details how Meta evolved from 6K GPU clusters with 128-512 GPU job sizes in 2022 to 100K+ GPU clusters in 2024-2025 with 16K GPU job sizes, driven by the explosive growth of Llama open-source LLMs. All platforms referenced are contributed to the OCP database with full specifications publicly available.

Prometheus incorporates two GPU platform designs: Clemente (based on NVIDIA GB300, housing 72 GPUs per rack with compute trays and ME switch trays alongside an Open Rack V3 High Power companion rack for PSUs, BBUs, and network switches) and Catalina (based on NVIDIA GV200, with two cross-rack cable-connected GPU racks of 36 GPUs each). Both designs use AALC (Air Assisted Liquid Cooling) racks. The cluster also includes storage platforms (Grand Canyon for HDD bulk storage, Olympic and Yosemite V3.5 Crater Lake for flash), compute platforms (Yosemite V3.5 Half Dome and Yosemite V4), and a network architecture built on DSF (Disaggregated Scheduled Fabric) with Broadcom Jericho3-AI and Ramon3 switches plus NSF (Non-Scheduled Fabric) based on Minipack 3 at 400G.