MetaRoCE: A New RDMA Transport Built for AI-Scale Ethernet
Meta open-sourced MetaRoCE, a new RDMA transport protocol built for million-GPU AI clusters on Ethernet.
“The fabric sees packets, but the NIC sees intent.”
Meta has designed and open-sourced MetaRoCE, a clean-sheet RDMA transport protocol purpose-built for AI training and inference at scale on commodity Ethernet. Unlike standard RoCE, MetaRoCE moves network intelligence to endpoint NICs rather than fabric switches, enabling native out-of-order packet delivery, packet spraying across paths, and per-path telemetry for large multi-data-center GPU clusters. The spec, reference implementation, and compliance suite are being released through the Open Compute Project, signaling Meta's push to make Ethernet the dominant fabric for AI infrastructure.