Meta has introduced MetaRoCE, a clean-sheet RDMA transport protocol designed specifically for AI workloads on commodity Ethernet. Unlike standard RoCE, which relies on lossless networks and Priority Flow Control (PFC), MetaRoCE treats the fabric as lossy and moves ordering, path selection, and recovery logic into the Network Interface Card (NIC).
- MetaRoCE enables out-of-order packet delivery with direct memory writes, eliminating reorder buffers and head-of-line blocking.
- The protocol supports native multipathing by using distinct UDP source ports for ECMP entropy, allowing the NIC to dynamically reroute traffic away from congested paths.
- It maintains approximately 86% throughput at 1% packet loss and continues delivering useful bandwidth even at 10% loss rates on a 64-node AMD GPU cluster.
- MetaRoCE requires only ECN marking and ECMP from switches, making it compatible with vendor clouds where operators cannot control switch configuration.
The specification, a compliance test suite, and a reference software implementation will be released through the Open Compute Project (OCP) at the 2026 OCP Global Summit in October.