Meta wants AI training to reach a million GPUs on ordinary Ethernet
Meta showed a network protocol meant to scale to million-GPU training and still run on different network topologies and different hardware.
At OCP Summit 2026, Meta showed MetaRoCE, an open RDMA transport layer. An OCP NIC 3.0 adapter was on display at its booth.
MetaRoCE adds features like packet spraying and congestion control to that transport. Meta's presentation listed reliability and congestion control that work per network path, and receiver rate hints (RCN) for fast incast convergence. It also tolerates packet loss and reordering, and a single queue pair (QP) can carry multiple streams over multiple network paths. It operates on commodity Ethernet using ECN and ECMP.
Before the summit, Meta demonstrated the protocol on the AMD Pensando NIC series. That NIC's programmability let Meta implement the protocol without dedicated hardware. ServeTheHome says this potentially lets Meta source from multiple vendors.
One small aside: the NIC on display used the SFF pull tab design, which ServeTheHome names as its favorite form factor.
Would an open network transport for giant AI clusters matter to you more than the chips themselves?
Archive video: Complexity Academy, CC BY 3.0 (Wikimedia Commons)
Video: ServeTheHome on YouTube