1、Xu WangYuvin Madhaya WeerasinghewithJinghan(Jeremy)YangStrategy for High-speed Cable Backplane Validation and Diagnostics in AI SystemsStrategy for High-speed Cable Backplane Validation and Diagnostics in AI SystemsXu WangYuvin Madhaya Weerasinghewith Jinghan(Jeremy)YangOCP SPECIAL FOCUS:ARTIFICIAL
2、INTELLIGENCE(AI)Topology:Single-hop switchedRack types:ORv3-HPR,etc.AI Systems and Scale-Up NetworkBandwidth per GPU in Tbps NPC to NPC cable lengths:2m metersCopper Cable Backplane OverviewPassive copper cable backplane is a key enabler of thescale-up network.oPower consumptionoCostoReliabilityoLow
3、er loss vs.PCB routingChallengesoDesign,validation,diagnostics,manufacturing,screeningoNeed for industry collaborationDC-Open IssueConnector mating DesignoDebrisoContact matingoX,Y,Z float mechanismoGuide pinsSerDes link quality TuningoPRBS(L2)oLink flappingoLong term traffic testingOur Experience w
4、ith Cable BackplanesChallengesLack of visibility/observabilityIssue isolationRoot causingTime consuming DOEsTolerance stack-upoExpected mating gapMating force and sequenceFloat and movementRetention and latchingThermal managementObservabilityCable bending and tensionCable Backplane Design Considerat
5、ionsConnector Mating Gap CheckCable Tension CheckDPPM per differential pairCable Backplane Quality MetricsReseat Success RateReseat Cycle LimitThere doesnt seem to be any well-established quality metric or failure rate with passive backplanes,and we propose the following:Cable Backplane ValidationPr
6、e-IntegrationPost-IntegrationMechanicalFunctionalMating gap inspectionManual re-seating testingTray/component backplane interface test(PRBS)FunctionalSignal integrity test vehicle PRBS testingPartial/full snake tests with packet trafficSerDes and link diagnostics data collectionS