服务器横向扩展互连技术 - UALink 超级节点和 CXL 内存池.pdf

1、VincentKong,Chief Ultra Link Architect-AlibabaServer Scale-Up Interconnect Technologies:UALink SuperNode and CXL Memory PoolServer Scale-Up Interconnect Technologies:UALink SuperNode and CXL Memory PoolVincentKong,Chief Ultra Link Architect-AlibabaOCP SPECIAL FOCUS:ARTIFICIAL INTELLIGENCE(AI)Cloud I

2、nfrastructure Needs Driven by ApplicationsNew application-driven deep collaboration and serverless,TCOAI-driven high performance GPU links and large memory requirementsCloudApplicationsGeneralComputingHeterogeneouscomputingAlibaba Server Scale Up SystemsLow latencyMemory semanticsLarge Pooling elast

3、icityData coherencyUltra-high bandwidth,extremely low latencyMemory semanticsMemory sharingThe heterogeneous is straight outALS（Accelerator Link System）CLS（CXL System)Scale Up Systems for CPUs and GPUsProtocolForm FactorTypical TrafficData PattenCommunication semanticsTypical BandwidthDomain SizeSca

4、le UpNvlink、UALinkGPU IOTP、EPHuge data size（Extremely latency-sensitive）Memory Semantics10TbpsSeveral RacksPCIe、CXLCPU IOMemory AccessCacheLine（Extremely latency-sensitive）Memory Semantics1TbpsSeveral RacksScale OutIB、UECPCIe CardDP、PPLarge data blockRDMA4800GbpsClusterCommon Need of Server System S

5、cale Up Fabricmemory semantics for GPU&CPULimited number of nodes,but ultra high performanceCacheLine coherent or IO coherent CXL is not only a memory hierarchy technology,it also enables tighter collaboration among multiple CPU nodes.Memory and storage are effectively tiered in Cloud Computing duri

6、ngpast 10 years,now SuperNodeServer is changing interconnect architectureParallel Algorithm Needb h 1 2 4 b h t 1 8EPEP DataData SizeSize ofof singlesingle OP:OP:TPTP DataData SizeSize ofof singlesingle OP:OP:Comparison of Data Transferred in EP/TP/PP/DP:Comparison of Data Transferred in EP/TP/PP/DP