Meta AI Unveils MetaRoCE for AI-Scale Ethernet Networking

Meta AI has introduced MetaRoCE, a novel RDMA transport protocol specifically engineered for AI-scale networking over commodity Ethernet. This development addresses the growing challenge where network performance, rather than compute power, is becoming a bottleneck for training and serving large frontier models. Collective operations, such as all-reduce and all-to-all, are crucial for synchronizing thousands of accelerators during model training, and any network latency can significantly impede the overall job completion time, leading to substantial compute capacity being stranded. MetaRoCE fundamentally rethinks the assumptions of standard RoCE (RDMA over Converged Ethernet). While standard RoCE relies on the network to deliver every frame in order, often utilizing Priority Flow Control (PFC) and discouraging packet spraying in large-scale networks, MetaRoCE treats the network fabric as inherently lossy. It shifts the responsibilities of ordering, path selection, and recovery to the Network Interface Card (NIC). This design aims to improve performance and resilience in complex, large-scale AI training environments. Meta is making the MetaRoCE specification, a reference software implementation optimized for DPDK (Data Plane Development Kit), and a compliance test suite available through the Open Compute Project (OCP). These artifacts are anticipated to be released, possibly in October 2026, with Meta potentially unveiling the specification, the DPDK-optimized software reference implementation, and its production compliance framework at the 2026 OCP Global Summit. Early hardware support for MetaRoCE has been demonstrated on AMD Pensando programmable NICs, with other vendors also developing implementations. This initiative represents a significant fabric architecture decision for AI infrastructure, moving beyond a simple procurement choice. Meta has experience scaling AI clusters to hundreds of thousands of GPUs across multiple data centers and regions, where the network is a critical path component for every training step. The limitations of standard RoCE, which expects ordered frame delivery and relies on PFC, have become apparent in these massive deployments. MetaRoCE's approach of pushing intelligence to the NIC, treating the fabric as lossy, and managing ordering and recovery at the endpoint is designed to overcome these limitations and unlock greater efficiency in AI networking.