Key insights
- OpenAI's partnership with major chipmakers to release the MRC protocol aims to improve AI training efficiency by eliminating bottlenecks. This could lead to faster AI development and deployment, benefiting companies like Nvidia, AMD, and Broadcom. The increased efficiency in AI model training may indirectly boost productivity and innovation across various sectors, positively influencing the US equity market.
Investing.com - OpenAI has partnered with AMD, Broadcom, Intel, Microsoft, and Nvidia to release Multipath Reliable Connection, a new open-standard protocol designed to eliminate bottlenecks in massive AI training clusters. The protocol was announced Tuesday through the Open Compute Project.
MRC extends RDMA over Converged Ethernet, an InfiniBand Trade Association standard that enables hardware-accelerated remote direct memory access among GPUs and CPUs. The protocol splits network interfaces into multiple smaller links to create parallel networks, allowing a single transfer to spread packets across hundreds of paths through the network.
The protocol is deployed across OpenAI’s largest Nvidia GB200 supercomputers used to train frontier models, including the company’s site with Oracle Cloud Infrastructure in Abilene, Texas, and in Microsoft’s Fairwater supercomputers. MRC has been used to train multiple OpenAI models, leveraging hardware from Nvidia and Broadcom.
MRC uses IPv6 Segment Routing to allow senders to directly specify the path each packet should take through the network, replacing dynamic routing protocols. The protocol detects network failures and routes around them on a microsecond timescale, compared to seconds or tens of seconds for conventional network fabrics.
OpenAI reports more than 900 million people use ChatGPT every week. During training of a recent frontier model, the company rebooted four tier-1 switches without coordinating with teams running training jobs in the cluster.
This article was generated with the support of AI and reviewed by an editor. For more information see our T&C.