OpenAI releases MRC, its network protocol for large-scale training clusters

OpenAI open-sources its MRC network protocol, co-developed with NVIDIA, Microsoft, and AMD to prevent GPU link failures in large AI training clusters.

OpenAI is open-sourcing the specifications for MRC (Multipath Reliable Connection), a network protocol developed over two years with AMD, Broadcom, Intel, Microsoft, and NVIDIA to ensure reliable communication between GPUs in AI training supercomputers. The specification has been contributed to the Open Compute Project, making it usable by the entire industry.

The problem addressed is familiar to any operator of a massive cluster: as the number of GPUs increases, link failures and intermittent congestions become inevitable, and a single incident can halt a synchronous training run involving tens of thousands of cards. MRC circumvents this fragility by distributing each transfer across hundreds of simultaneous paths and isolating a faulty route within microseconds, whereas a classical architecture could take several seconds to reconfigure.

The protocol integrates with the latest 800 Gb/s network interfaces and allows for reducing the topology of large clusters to two tiers of switches instead of three or four, resulting in lower power consumption and fewer components prone to failure. MRC is already running on OpenAI's largest NVIDIA GB200 supercomputers, including the Stargate site in Abilene, Texas, operated with Oracle Cloud Infrastructure, and Microsoft's Fairwater infrastructure, and has been used to train several in-house models.

The open-sourcing of the standard is part of a deliberate strategy: rather than keeping this component proprietary, OpenAI is betting on industry-wide standardization to advance the hardware ecosystem on which future frontier models depend.