AMD hands AI networking protocol to Open Compute Project

AMD has contributed the Multipath Reliable Connection networking protocol to the Open Compute Project, working with OpenAI, Microsoft and other industry groups on the effort.

The move adds a new AI networking protocol to the Open Compute Project as companies race to build larger computing clusters for model training. The protocol is intended for systems where hundreds of thousands of GPUs must remain synchronised, exchange data continuously and recover quickly from network disruptions.

MRC is designed for large-scale AI training environments, where traditional single-path networking can create bottlenecks. Instead of sending traffic over one route, it distributes packets across multiple paths at the same time to reduce congestion and limit latency swings.

The approach is intended to keep tightly coupled AI training jobs running during failures. MRC can reroute traffic in near real time, avoiding the slower recovery processes associated with conventional network methods.

AMD presented the protocol as part of a broader push to make AI infrastructure more open and programmable. Contributing it to the Open Compute Project could support wider industry adoption by making the specification available to the broader ecosystem.

Industry pressure

Networking has become a bigger issue for AI developers as clusters grow in size and complexity. In these environments, delays in communication between processors can leave expensive compute resources underused, making network efficiency both a commercial and technical concern.

According to AMD, MRC reflects the needs of cloud providers, enterprises, research groups and sovereign AI projects building increasingly large clusters. These users need systems that perform reliably under normal operating conditions, not just under ideal tests.

AMD helped write the MRC specification and contributed congestion control technology intended to improve performance under real-world workloads. It has also implemented and deployed MRC with AMD networking technology in test clusters with a leading cloud provider, though it did not identify the provider.

That testing suggests the protocol has moved beyond a draft standard into live operational environments. AMD described MRC as an evolution of a pre-standard implementation of an improved RoCEv2 transport protocol it had already developed.

Hardware link

AMD also tied the protocol closely to its own networking products, particularly its Pensando Pollara 400 AI NIC. The company said the card's hardware and software programmability helped validate early versions of the technology and implement MRC on a 400G network interface card.

AMD added that this work could support a transition to its Pensando Vulcano 800G AI NIC, which supports the same transport protocol. That connects the standardisation effort to its broader attempt to strengthen its position in AI infrastructure beyond processors and accelerators.

Open standards bodies have become an important battleground in data centre technology as suppliers seek to influence the technical foundations of next-generation systems. By contributing MRC to the Open Compute Project, AMD and its collaborators are trying to shape how Ethernet-based AI networks develop at a time when demand for large training systems is rising sharply.

The companies involved include OpenAI, Broadcom, Intel and Microsoft alongside AMD. Their participation underlines how networking is becoming a shared challenge across the AI supply chain, from cloud operators to chipmakers and model developers.

Krishna Doddapaneni, corporate vice president of engineering, NTSG, AMD, highlighted the commercial importance of the network layer in AI systems.

"As GPUs and CPUs continue to drive compute, real bottleneck in scaling AI is the network. AMD, alongside OpenAI and Microsoft announced MRC, marking a major step forward for the industry. The programmability from AMD enables us to rapidly turn innovations like this into real-world performance at scale, where consistent, resilient throughput matters more than theoretical peak bandwidth," Doddapaneni said.