OpenAI has teamed up with AMD, Broadcom, Intel, Microsoft, and NVIDIA to develop MRC, an open source network protocol that sends data across hundreds of paths simultaneously between GPUs. Instead of three or four switch layers, MRC needs only two to connect over 100,000 GPUs, cutting both power and costs. The protocol is already running on OpenAI's Stargate supercomputer.
The article OpenAI built a networking protocol with AMD, Broadcom, Intel, Microsoft, and NVIDIA to fix AI supercomputer bottlenecks appeared first on The Decoder.
