Meta AI Unveils MetaRoCE for AI-Scale Ethernet Networks

Training frontier AI models requires thousands of accelerators to exchange and synchronize data at high speed, making network congestion and latency major constraints on cluster performance. Meta AI designed MetaRoCE as a clean-sheet RDMA transport for commodity Ethernet, aiming to deliver the predictable, efficient data movement needed by large-scale AI workloads without relying on specialized networking infrastructure.

Meta AI has now detailed MetaRoCE, which shifts packet ordering and multipath path selection to the network interface card, or NIC. The design is intended to use multiple network paths more efficiently and reduce synchronization bottlenecks as accelerator clusters scale. Meta plans to publish the protocol specification and an open-source reference implementation through the Open Compute Project, though it has not announced a firm release date.