Meta's MetaRoCE reinvents RDMA transport for million-GPU Ethernet clusters
Meta's clean-sheet RDMA transport tosses out the core assumption standard RoCE is built on: that the fabric must deliver frames in order. MetaRoCE moves intelligence to the NIC, sprays packets across many paths, and treats out-of-order arrival as normal so every GPU transfer lands straight in memory with no reorder buffer or head-of-line blocking. The spec, a reference implementation, and a compliance suite are now out under the Open Compute Project. It is the clearest industry signal yet that Ethernet can carry AI at scale without the fabric strangling compute.