OpenAI has introduced Multipath Reliable Connection (MRC), a novel supercomputer networking protocol designed to improve performance and resilience in large-scale AI training clusters. Developed over two years in collaboration with AMD, Broadcom, Intel, Microsoft, and NVIDIA, MRC has been released through the Open Compute Project to benefit the broader industry. Frontier model training relies heavily on reliable networks to move data quickly between GPUs, but network congestion and device failures often cause delays that idle GPUs and disrupt training jobs. MRC addresses these challenges by extending RDMA over Converged Ethernet and incorporating SRv6-based source routing. Instead of assigning a data transfer to a single path, MRC sprays packets across hundreds of paths through multi-plane networks, virtually eliminating core congestion and routing around failures in microseconds. This approach allows networks to connect over a hundred thousand GPUs using only two tiers of switches, reducing power consumption, component count, and overall cost. MRC is already deployed across OpenAI's largest NVIDIA GB200 supercomputers, including sites with Oracle Cloud Infrastructure in Abilene, Texas, and Microsoft's Fairwater supercomputers, and has been used to train multiple OpenAI models. Alongside the release, OpenAI and its partners co-authored a technical paper detailing their experiences with the new protocol.
Unlocking large scale AI training networks with MRC (Multipath Reliable Connection)
OpenAI has introduced Multipath Reliable Connection (MRC), a new supercomputer networking protocol developed in partnership with AMD, Broadcom, Intel, Microsoft, and NVIDIA. Released via the Open Compute Project, MRC aims to improve resilience and performance in large-scale AI training clusters by reducing network complexity, congestion, and failure impacts.
Why it matters
Efficient networking is critical for scaling large AI training clusters, where network congestion and hardware failures can idle expensive GPUs and stall training jobs. Shared infrastructure standards like MRC can help the industry scale AI systems more reliably and cost-effectively.