MRC Protocol: Supercomputer networking to accelerate large scale AI training (openai.com)

OpenAI has introduced the Multipath Reliable Connection (MRC) protocol in collaboration with AMD, Broadcom, Intel, Microsoft, and NVIDIA to enhance GPU networking performance for large-scale AI model training on supercomputers. MRC facilitates rapid data transfer between GPUs by allowing a single data transfer to be split across hundreds of paths, which effectively mitigates the impact of network congestion and failures—issues that have become more pronounced as the scale of AI training jobs increases. This innovative approach enhances efficiency, minimizes downtime, and supports the training of increasingly sophisticated AI models. The significance of MRC lies in its potential to streamline AI infrastructure by establishing a standardized networking protocol that can be adopted across the tech industry. By publishing the MRC specification as part of the Open Compute Project, OpenAI aims to foster collaboration and innovation among partners, reducing complexities associated with networking at scale. Key technical implications include MRC's use of static source routing through SRv6, which simplifies network management and reduces the likelihood of failures interrupting training jobs. As it operates efficiently even during link failures, MRC significantly lowers the costs and resource consumption associated with building robust AI training systems, allowing for faster delivery of advanced AI models.