MetaRoCE: A New RDMA Transport Built for AI-Scale Ethernet

Meta introduces MetaRoCE, a clean-sheet RDMA transport protocol designed for AI workloads on commodity Ethernet at million-GPU scale. Unlike standard RoCE, which relies on in-order delivery and PFC-based losslessness, MetaRoCE moves intelligence to the NIC, embracing native out-of-order delivery, multipathing with per-path telemetry, loss tolerance without pause frames, and dual-sided congestion control. Tested on a 64-node AMD GPU cluster with Pensando NICs, it showed higher throughput and lower flow completion times than RoCEv2, maintaining ~86% throughput at 1% packet loss and degrading gracefully even at 10% loss. Meta is contributing the specification, a compliance suite, and a software reference implementation (libsoftmetaroce) to the Open Compute Project, with a full release planned at the 2026 OCP Global Summit in October.

Source: https://engineering.fb.com/2026/08/24/networking-traffic/metaroce-rdma-transport-ai-ethernet. 8 Sync News only summarizes and links out; content copyright belongs to the authors and original sources.