MetaRoCE: A New RDMA Transport Built for AI-Scale Ethernet

Meta's MetaRoCE networking protocol puts responsibility for handling disrupted transfers into the network adapters connecting AI accelerators. Its engineering account describes spreading packets across multiple paths, writing arriving data directly to its destination in memory and retransmitting missing packets. A congested path can therefore be handled separately from the rest of the connection.

This matters when distributed training must wait for data exchanges to finish. In a reported 64-node GPU-cluster test, the protocol retained about 86% throughput at 1% packet loss. That provides a concrete, bounded test of the design's ability to keep communication useful when the network drops data.