OpenAI releases MRC networking specification for large AI clusters

OpenAI released the Multipath Reliable Connection specification through the Open Compute Project on May 5. MRC extends RoCE networking for large GPU clusters, spreading a transfer across many paths and using SRv6 source routing to move around failures. OpenAI says the protocol is deployed on its largest NVIDIA GB200 supercomputers and has been used in model training.

Many smaller planes replace one large path

The published design splits an 800 Gb/s interface into multiple lower-rate links connected to different switches. That creates parallel network planes and more physical path diversity. MRC then sprays packets across those paths instead of pinning each transfer to one route. The receiver reorders packets, while the sender adapts to congestion and failures. Static source routes reduce dependence on a distributed routing control plane reacting during a training step.

The rationale comes from synchronous training: one delayed transfer can hold GPUs idle across a collective operation. At cluster scale, link and switch faults are routine background conditions rather than exceptional events. The protocol therefore optimizes predictable progress during partial failure, not only peak link speed.

Publication makes interoperability the next test

An open specification gives NIC, switch, and cluster operators a common artifact to inspect, but deployability depends on hardware support and operational tooling. Teams evaluating MRC should examine receiver reordering resources, congestion fairness, path selection, telemetry, failure injection, and compatibility with their collective libraries and topology.

The architectural choice also shifts complexity. A simpler network control plane is paired with smarter endpoints and explicit route construction. That trade can be attractive when training traffic is regular and high-volume, but it needs validation under mixed workloads, partial upgrades, and degraded links. The specification is a starting point for that work, not evidence that every AI cluster should adopt the same fabric.