OpenAI collaborates with five industry giants to set protocol standards, aiming at the million-GPU cluster bottleneck

OpenAI announced on Wednesday, in conjunction with AMD, Broadcom, Intel, Microsoft, and Nvidia, the release of an open networking protocol called "Multipath Reliable Connection" (MRC), aimed at addressing network congestion and failure issues in AI training clusters at the scale of millions of GPUs.

The protocol has been fully deployed on all of OpenAI's cutting-edge model training supercomputers.

The Network Bottleneck in the Million-GPU Era Spurs the Creation of a New Protocol

MRC is a new type of network protocol based on the RDMA (Remote Direct Memory Access) architecture, utilizing packet injection technology to allow concurrent transmission of traffic over multiple network paths, thereby fully utilizing the cluster bandwidth.

In terms of fault handling, MRC supports microsecond-level link switching. When a single network link fails, traffic can be rerouted to a backup path in an extremely short time, effectively reducing the compute power wasted by GPUs waiting for network recovery.

OpenAI stated that as AI training scales to a magnitude of hundreds of thousands or even millions of GPUs, traditional network protocols face significant congestion bottlenecks, and MRC is a systemic solution to this problem.

MRC has currently been deployed in two of OpenAI's core training infrastructures: the supercomputing site in Abilene, Texas, U.S., built on Oracle Cloud, and Microsoft's Fairwater supercomputing cluster.

The Battle of Open Protocols: Challenging InfiniBand's Closed Ecosystem

The industry interprets OpenAI's decision to release MRC in the form of an open protocol and to collaborate with major players in the chip and cloud computing industries as a clear standard-setting action.

Should MRC gain widespread adoption, it would pose direct competitive pressure on traditional closed high-speed interconnect solutions such as Mellanox InfiniBand. Nvidia's presence in the list of MRC collaborators indicates that while maintaining its current ecosystem, it is also seeking a position in emerging open standards.

Analysts have pointed out that the competition in AI infrastructure has extended from chip computation to the cluster network layer, and mastering the leading role in interconnect standards will have a profound impact on the construction pattern of future hyper-scale data centers.