Meta unveiled MetaRoCE this week, a transport protocol built from scratch for AI traffic on commodity Ethernet. The old protocol, standard RoCE, assumes the network delivers every frame in order and leans on PFC. MetaRoCE inverts that logic: the fabric is treated as lossy, and the NIC owns ordering, path selection, and recovery.
The stakes are high at Meta’s scale. Collective operations such as all-reduce and all-to-all synchronize hundreds of thousands of GPUs across data centers, and one slow transfer paces the entire training job.
Loss tolerance is the headline result. Throughput holds at about 86% when 1% of packets drop, and the fabric still carries useful bandwidth at 10% loss. The tests used a 64-node AMD GPU cluster with RCCL collectives. MetaRoCE beat RoCEv2 on throughput and finished flows faster. Multi-plane tests scaled linearly with plane count, and simulated plane failures saw traffic redistribute without operator input.
Meta is handing the work to the Open Compute Project: a specification, a compliance suite for vendors, and libsoftmetaroce, a reference implementation that serves as the behavioral model for silicon development. The artifacts are expected around the October OCP Global Summit, and the project extends OCP’s Ethernet Scalable Unified Network initiative. The first hardware proof ran on AMD Pensando programmable NICs, with other vendors’ implementations in progress.
