OpenAI Teams Up with Tech Giants to Launch MRC Protocol, Addressing Network Bottlenecks in AI Training

Large-scale AI model training often leads to GPU idling due to network delays and failures, with problems worsening as cluster size increases. Recently, OpenAI has joined hands with AMD, Broadcom, Intel, Microsoft, and NVIDIA to launch the Multipath Reliable Connection (MRC) protocol, open-sourcing it via the Open Compute Project (OCP).

MRC uses multi-plane design to split interfaces and reduce architectural complexity, adaptive packet spraying to distribute data, and SRv6 source routing to simplify control, cutting failure recovery time from seconds to microseconds. Currently applied in scenarios like NVIDIA GB200 supercomputers, it can bypass failures without interrupting training.