OpenAI unveils MRC, a network protocol for supercomputers with millions of GPUs

OpenAI, in collaboration with AMD, Broadcom, Intel, Microsoft and NVIDIA, has introduced MRC (Multipath Reliable Connection), a new open networking protocol for large AI clusters. It targets a key problem: how to build a supercomputer with hundreds of thousands of GPUs when network reliability becomes the bottleneck.

How MRC works

MRC spreads data packets across hundreds of network paths at the same time. If one path fails, data flows along alternative routes without losing speed. Recovery from failures takes microseconds, so fast that neural network training barely notices the disruption.

Traditional network protocols pick one primary path and switch to a backup only if the primary breaks. MRC works on a fundamentally different principle: it monitors hundreds of potential paths in real time and dynamically distributes load across the healthy links. It is like the difference between a single road with a detour and a web of side streets where cars can take any free route.

Practical advantages

MRC's main achievement is a simpler supercomputer architecture. Clusters with 100,000+ GPUs previously required a three-tier hierarchy of Ethernet switches. That is expensive, complicated to install, and demands a lot of power for cooling. MRC gets by with just two tiers, which radically simplifies the design and cuts hardware costs.

• Less network hardware, which simplifies installation and maintenance
• Lower network latency thanks to more direct paths between GPUs
• Lower power consumption for switch cooling
• Better scaling: the architecture works up to millions of GPUs

An open standard for the ecosystem

MRC is not a closed solution owned by one company. OpenAI chose an open approach and brought in the major networking hardware makers: AMD, Broadcom, Intel, Microsoft and NVIDIA. That means other companies, cloud providers and research centers can deploy MRC in their own supercomputers.

An open standard matters precisely because at this scale even small improvements in network reliability and efficiency affect the cost of training models and the pace of the entire AI industry.

What it means

MRC is a response to the challenge of scale. As AI models grow, so do the demands on computing infrastructure. A network architecture that works for a cluster of 10,000 GPUs can become a bottleneck at 500,000 GPUs. MRC makes it possible to build even larger supercomputers without radically rethinking the architecture.

For the industry, this means cheaper model training and faster adoption of innovations.