OpenAI and Chipmakers Launch MRC Protocol for AI Clusters

OpenAI Unveils New Network Protocol with Tech Giants

OpenAI, in collaboration with industry leaders including AMD, Broadcom, Intel, Microsoft, and Nvidia, has announced the Multipath Reliable Connection or MRC protocol. This new open-standard was released through the Open Compute Project to address and eliminate network bottlenecks in large-scale AI training clusters.

Protocol Overview and Deployment

MRC enhances RDMA over Converged Ethernet by splitting network interfaces into multiple smaller links. This creates parallel networks, allowing a single data transfer to be distributed across hundreds of paths. The protocol is already deployed in OpenAI's largest Nvidia GB200 supercomputers used for training its frontier models, including facilities with Oracle Cloud Infrastructure and Microsoft’s supercomputers.

Technical and Market Impact

Utilizing IPv6 Segment Routing, MRC allows senders to specify packet paths directly, which improves network efficiency. It can detect and route around network failures on a microsecond timescale, a significant improvement over the seconds required by conventional fabrics. This advancement is crucial for maintaining performance during the intensive training of models that support services like ChatGPT, used by over 900 million people weekly.

Conclusion

The introduction of the MRC protocol marks a significant step toward standardizing high-performance networking for AI infrastructure. This collaboration between OpenAI and major chipmakers is expected to accelerate the development of more powerful and efficient AI models by creating a more resilient and faster training environment.

FAQ

Q: What is the Multipath Reliable Connection (MRC) protocol?
A: It is a new open-standard network protocol designed by OpenAI and its partners to eliminate bottlenecks and improve efficiency in massive AI training clusters.

Q: Which companies partnered with OpenAI on the MRC protocol?
A: The partnership includes AMD, Broadcom, Intel, Microsoft, and Nvidia.

Source: Investing.com