释放人工智能规模化发展的潜力：UALink与OCP的合作.pdf

1、Kurtis Bowman,UALink Consortium Board Chair,AMDUnleashing AI Scale-Up:UALink and OCP CollaborationUnleashing AI Scale-Up:UALink and OCP CollaborationKurtis Bowman,UALink Consortium Board Chair,AMDOCP SPECIAL FOCUS:ARTIFICIAL INTELLIGENCE(AI)Advancing AI Across Data CentersAI models continue to grow,

2、they require more compute and memory to efficiently execute training and inference on large modelsThe industry needs an open solution that enables efficient distribution of models across many accelerators within a podLarge inference models will require scale-up of 10s 100s of accelerators in podsLar

3、ge training models will require scale-up and scale-out from 100s 10,000s of accelerators by connecting multiple podsUltra Accelerator Link TimelineMay 2024Promoter Group Press ReleaseOctober 2024UALinkMembership Forms Posted To WebsiteApril 2025UALink 200G 1.0 Specification PublishedAugust 2025UALin

4、k 200G 1.0 Specification Receives FMS Most Innovative Technology AwardMay 2025UALink Consortium reaches 100 membersSeptember 2025UALink 128G DL-PL 1.0 Specification PublishedUALink Creates the Scale-up PodHigh performanceUp to 800Gbps per Port,scalable ports per accelerator,Up to 1,024 acceleratorsL

5、ow latencyOptimized protocol,transaction,link&physicalLow powerThe simplified UALink stack leads to lower power solutionsLow die areaOptimized data layer and transaction layer saves significant die area UALink delivers optimized scale-upsolutions with single-tier switchingScale-Out NetworkThe UALink

6、 interconnect enables Accelerator-to-Accelerator communicationThe initial focus is sharing memory among acceleratorsDirect load,store,and atomic operations between accelerators(i.e.GPUs)Low latency,high bandwidth fabric for 100s of accelerators in a pod(up to 1K)Simple load/store/atomics semantics w