Meta shares its latest networking infrastructure innovations for gigawatt-scale AI, including the IcePack liquid-cooled network rack, Flattened Planar Fabric (FPF) for improved scale-out, and the MetaRoCE RDMA transport specification. These developments aim to optimize cluster-wide performance by reducing bottlenecks in collective communication across hundreds of thousands of accelerators.
Key points
Meta's Flattened Planar Fabric (FPF) reduces network tiers to a 2-tier Clos architecture to improve scale-out efficiency.
MetaRoCE provides an open-source RDMA transport protocol specifically optimized for the high-performance demands of AI collective operations.
The transition to liquid-cooled racks with 102.4T switch blades enables higher density for gigawatt-scale superclusters.
Ethernet is being utilized for scale-up interconnects to support massive accelerator clusters, moving away from proprietary protocols.