Meta Releases MetaRoCE, a Clean-Sheet RDMA Transport for AI-Scale Ethernet

Meta released MetaRoCE, a clean-sheet RDMA (Remote Direct Memory Access) transport protocol purpose-built for AI workloads running on commodity Ethernet. The goal is to move data directly between GPU memories across a data center fabric without burning compute cycles on the transfer—a critical bottleneck as training scales to hundreds of thousands of accelerators.

Rather than adapt legacy RoCE (RDMA over Converged Ethernet), which was designed for general data-center traffic, Meta rebuilt the transport from scratch around the traffic patterns of large-scale AI training. The company is publishing the full specification, a reference software implementation, and a compliance test suite—an open approach intended to let hardware and network vendors build interoperable implementations.

The release aligns Meta with an industry-wide move toward Ethernet-based AI networking as an alternative to proprietary InfiniBand, a fabric NVIDIA dominates. NVIDIA's own Spectrum-X Ethernet push, detailed the same week, and the Ultra Ethernet Consortium reflect the same competitive dynamic: hyperscalers want open, commodity-Ethernet fabrics they can source from multiple vendors rather than single-supplier interconnects.

Open-sourcing a transport spec is a bid to shape a standard, but adoption depends on NIC and switch vendors implementing it and on real-world performance matching InfiniBand for the largest runs. MetaRoCE pairs naturally with Meta's MTIA 300 accelerator and its built-in NIC chiplets, hinting at a coherent in-house networking-plus-silicon strategy. Readers should watch which vendors adopt MetaRoCE, how it benchmarks against Spectrum-X and InfiniBand, and whether it converges with or competes against Ultra Ethernet.