The key factor determining the performance of mega-AI models is shifting from 'GPU computing power' to 'network efficiency'. This is because in an AI training environment where hundreds of thousands of GPUs synchronize and exchange data in real time, even a single delay or failure can reduce overall throughput.

Under these circumstances, AMD is attracting attention from the industry by unveiling a new AI network standard, 'MRC (Multipath Reliable Connection),' together with OpenAI, Microsoft, and others.

AMD contributed MRC to the Open Compute Project (OCP) and opened it up for use across the ecosystem.

While existing single-path based networks have shown limitations in handling large-scale AI traffic, MRC reduces congestion and minimizes latency variation by simultaneously distributing and transmitting packets across multiple paths.

In addition, it is characterized by a design that readjusts the path in real-time in the event of a failure, allowing the network to function like a 'shock absorber'.

AMD stated that it went beyond simply participating in the establishment of the standard and implemented and verified MRC in actual cloud provider test clusters.

This means that MRC is not a theoretical proposal, but a technology whose performance has been proven in actual massive AI training environments.

Krishna Dodapaneni (CVP), head of AMD’s Networking division, emphasized, “The real bottleneck in AI scaling is the network,” adding that “AMD’s programmable networking technology quickly brings innovation into practice.”

In particular, AMD has already implemented MRC-based technology in the Pensando Pollara 400 AI NIC and plans to naturally extend this to the next-generation 800G 'Vulcano' AI NIC. The programmable architecture across both hardware and software is considered a differentiating factor for AMD compared to its competitors.

AI infrastructure performance is no longer defined by 'theoretical maximum bandwidth,' but by how stably GPUs can be kept at near 100% utilization in a real-world environment.

In response to these demands, MRC is expected to establish itself as a core technology that enhances the efficiency and reliability of large-scale AI clusters.

AMD announced its goal to expand the MRC ecosystem together with OpenAI, Intel, Broadcom, and others, and to develop AI networking into an open, standards-based infrastructure.