large-scale AI training environments<\/a>\u00a0where traditional single-path networking models\u00a0struggle.\u00a0These workloads require continuous, high-speed communication, and even brief disruptions can\u00a0impact\u00a0overall system progress.<\/p>\nInstead of sending traffic along a single path, MRC distributes packets across multiple paths simultaneously. This reduces\u00a0congestion\u00a0hotspots and limits latency variation that can slow synchronized training. When failures inevitably occur, MRC adapts quickly and allows traffic to reroute in near real-time, avoiding the delays associated with traditional network recovery.<\/p>\n
In practical terms, MRC helps turn the network into a shock absorber for AI infrastructure. Instead of forcing every event to become a disruption, MRC gives the network a way to adapt locally and\u00a0quickly\u00a0so workloads can continue making progress.\u00a0That matters because performance at\u00a0AI\u00a0scale is not defined by peak bandwidth alone. It is defined by how much useful accelerator capacity\u00a0remains\u00a0productive under real-world conditions.<\/p>\n
AMD Contributions: From\u00a0Development\u00a0to\u00a0Deployment<\/p>\n
AMD played a formative role in shaping how MRC works today.\u00a0AMD co-led authorship\u00a0the specification that defines next-generation AI networking and\u00a0contributed\u00a0advanced congestion control technology to improve performance under real-world conditions.<\/p>\n
More importantly, this\u00a0isn\u2019t\u00a0theoretical. AMD has implemented and deployed MRC, combined with AMD networking technology, at scale in\u00a0test clusters\u00a0with a leading cloud provider.\u00a0This validation means the design reflects how networks actually perform under sustained AI workloads.<\/p>\n
\u201cAs GPUs and CPUs continue to drive compute, the real bottleneck in scaling AI is the network. Today\u2019s MRC announcement from OpenAI marks a major step forward for the industry. AMD\u2019s programmability enables us to rapidly turn innovations like this into real-world performance at scale, where consistent, resilient throughput matters more than theoretical peak bandwidth.\u201d – Krishna Doddapaneni, CVP, Engineering, NTSG, AMD<\/p>\n
Programmability\u00a0remains\u00a0a key differentiator for AMD, as one of the only networking solutions that combines full hardware and software programmability with proven deployments, allowing networks to adapt as workloads evolve.\u00a0\u00a0Before the development of the MRC specification, AMD had a pre-standard implementation of an improved RoCEv2 transport protocol, which\u00a0evolved\u00a0into the MRC standard\u00a0of\u00a0today. This was due to the open programmability of the AMD\u00a0Pensando\u2122 Pollara 400 AI NIC, and that programmability contributed to the flexibility in obtaining early validation. As AMD being one of the first and\u00a0only companies to deploy MRC on a 400G NIC, we can accelerate a seamless transition to our AMD\u00a0Pensando\u00a0\u201cVulcano\u201d 800G AI NIC, which also supports the MRC transport protocol.<\/p>\n
This combination of\u00a0a\u00a0defined specification,\u00a0contributed technology, and implementation\u00a0in testing\u00a0positions AMD at the forefront of deploying MRC in real-world AI infrastructure.<\/p>\n
Redefining Performance for AI Infrastructure<\/p>\n
For AI at scale, performance is defined by how systems behave under real conditions, not peak bandwidth. Consistent throughput, effective congestion handling, and quick recovery from failures, while keeping GPUs synchronized and productive is\u00a0what\u2019s\u00a0optimal\u00a0to power AI networking at scale. MRC can improve model efficiency and helps make the networking protocols connecting large-scale AI training across large GPU clusters\u00a0highly reliable.<\/p>\n
By helping define and contribute to MRC, AMD is advancing AI networking from concept to practical, production-ready infrastructure.\u00a0<\/p>\n","protected":false},"excerpt":{"rendered":"What does it take to power the world\u2019s most demanding AI models, like those behind ChatGPT? At the…\n","protected":false},"author":2,"featured_media":29820,"comment_status":"","ping_status":"","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[7],"tags":[13248,5977,19510,157,19511],"class_list":{"0":"post-29819","1":"post","2":"type-post","3":"status-publish","4":"format-standard","5":"has-post-thumbnail","7":"category-openai","8":"tag-ai-intelligent-systems","9":"tag-corporate","10":"tag-instinct-gpus","11":"tag-openai","12":"tag-pensando-network-infrastructure"},"_links":{"self":[{"href":"https:\/\/www.europesays.com\/ai\/wp-json\/wp\/v2\/posts\/29819","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.europesays.com\/ai\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.europesays.com\/ai\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.europesays.com\/ai\/wp-json\/wp\/v2\/users\/2"}],"replies":[{"embeddable":true,"href":"https:\/\/www.europesays.com\/ai\/wp-json\/wp\/v2\/comments?post=29819"}],"version-history":[{"count":0,"href":"https:\/\/www.europesays.com\/ai\/wp-json\/wp\/v2\/posts\/29819\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.europesays.com\/ai\/wp-json\/wp\/v2\/media\/29820"}],"wp:attachment":[{"href":"https:\/\/www.europesays.com\/ai\/wp-json\/wp\/v2\/media?parent=29819"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.europesays.com\/ai\/wp-json\/wp\/v2\/categories?post=29819"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.europesays.com\/ai\/wp-json\/wp\/v2\/tags?post=29819"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}