UALink 概述.pdf

1、UALink Deep DiveUALink Deep DiveUALink ConsortiumUALinkDeep DiveKurtis BowmanBoard Chair,UALink ConsortiumAI CLUSTERSExplosive Growth of AI/HPCExplosive Growth of AI/HPCUnprecedented demand for compute and memory.Interconnects as the BottleneckInterconnects as the BottleneckTraditional and proprieta

2、ry solutions hinder scale and efficiency.The Need for Open StandardsThe Need for Open StandardsDriving choice,reducing costs,fostering innovation.The AI Infrastructure ImperativeCablingCablingConnectorsConnectorsReRe-TimersTimersManagement Management SoftwareSoftwareProtocol InterfaceProtocol Interf

3、aceTransaction LayerTransaction LayerData Link LayerData Link LayerIndustry Standard Ethernet PHYIndustry Standard Ethernet PHYEngineered to ScaleAnyAny CPUCPUAnyAny AcceleratorAcceleratorAnyAny SwitchSwitchThe Open Protocol for AI ScaleThe Open Protocol for AI Scale-Up NetworkingUp Networking100+Co

4、nsortium Members100+Consortium MembersOpen Standard Drives Innovation&ChoiceOpen Standard Drives Innovation&ChoiceUALink Creates the Scale-Up PodHigh High performanceperformanceUp to 800Gbps per Port,scalable ports per accelerator,Up to 1,024 acceleratorsLow latencyLow latencyOptimized protocol,tran

5、saction,link&physicalLow powerLow powerThe simplified UALink stack leads to lower power solutionsLow die areaLow die areaOptimized data layer and transaction layer saves significant die area UALink 1.0 focus is to deliver optimized scale-up solutions with single tier switchingUALink Stack Features&G

6、oalsUALinkPHY LayerIEEE P802.3dj Layer 1 Standard FEC Lower latency via 1-way and 2-way code word interleave Minor tweaks for 680-Byte FLIT code word alignmentFEC StatisticsPhysical Layer(incl.FEC)200Gbps/lane&100Gbps/lanePMAUALink TL/DL Features&Goals Fixed Payloads(64B/640B)Virtual Channels Link L