重塑背板：为什么人工智能需要积极主动的方法.pdf

1、Matthew Ho,TE ConnectivityChris Blackburn,Astera LabsReinventing the Backplane:Why AI Demands an Active ApproachReinventing the Backplane:Why AI Demands an Active ApproachMatthew Ho,TE ConnectivityChris Blackburn,Astera LabsOCP SPECIAL FOCUS:ARTIFICIAL INTELLIGENCE(AI)“AI Infrastructure 2.0”all serv

2、ers in a rack function as a unified computing platform rather than collections of individual serversCluster sizes of 100k+acceleratorsCable backplane architectures are status quo for Scale Up network fabricsHigh speed:224Gbps per lane 448GbpsHigh faceplate density:1.5 2x more dense than pluggable IO

3、Blind mating system chassisFixed fabric topologyTrends in AI System ArchitecturesCable Backplane CartridgeComputeComputeScale-Up SwitchesImpact of Trends in AI System Architectures#of accelerators in a Scale Up fabric#of XPU to XPU connections within a rack#of XPU to XPU connections between racksLat

4、ency between XPUsConnectivity bandwidth between XPUsScale Up protocol speedCable reachCable AWGScale Up ProtocolsDespite differences amongst the scale protocols,the trends are consistent!Rack interconnect densityo10k+connections per rackoTrends pushing towards wider racksoLimits to extending into ho

5、t aisleCable characteristicsoLarger cables enable reachoNarrower cables improve minimum bend radius and cartridge depthCable Backplane Architectural Challenges16.5mm13.5mmArchitecture OverviewLong reaches are retimed mid-channelShort reaches remain passiveTradeoffs and considerationsPower impact Cab

6、le Bundle SizeCartridge SizeServiceabilityThermals/CoolingPort ReconfigurabilityActive Cable Backplane SolutionDemo Overview2m 40OU tall cartridgeRepresents 30 compute and 10 switch sledsLive retimed PCIe Gen6 channelLive passive PCIe Gen6 channelServiceable r