OpenAI

OpenAI and Partners Release MRC Protocol for AI Supercomputer Networking


Executive Summary

OpenAI, in collaboration with AMD, Broadcom, Intel, Microsoft, and NVIDIA, has developed and released the Multipath Reliable Connection (MRC) protocol. Delivered as a specification through the Open Compute Project (OCP), MRC is a new networking protocol designed to improve the performance and resilience of large-scale AI training clusters. The protocol addresses key challenges like network congestion and link failures by spreading data transfers across hundreds of paths, enabling microsecond-level rerouting around failures and simplifying network design.

Key Takeaways

* Protocol Name: Multipath Reliable Connection (MRC).

* Core Function: MRC sprays packets from a single data transfer across hundreds of network paths simultaneously, which improves load balancing and fault tolerance.

* Microsecond Fault Tolerance: The protocol can detect network failures or congestion and reroute traffic in microseconds, minimizing interruptions to synchronous training jobs that are highly sensitive to delays.

* Simplified Network Design: MRC utilizes IPv6 Segment Routing (SRv6), allowing the sender to specify the exact path a packet travels. This eliminates the need for complex dynamic routing protocols like BGP, reducing a potential source of failure.

* Enables Multi-plane Topologies: The protocol is designed for multi-plane networks where a single high-speed interface is split into multiple lower-speed parallel networks. This increases path diversity, lowers costs, and reduces power consumption.

* Availability: The MRC specification is now available through the Open Compute Project (OCP) for industry-wide adoption. It is already deployed in OpenAI's largest NVIDIA GB200 supercomputers.

Strategic Importance

By open-sourcing MRC, OpenAI is standardizing a critical component of the AI infrastructure stack, aiming to unblock a key bottleneck in scaling AI models and fostering a broader ecosystem for building more efficient and resilient supercomputers.

Original article