利用 MGX 机架和数字孪生技术将任何数据中心改造为人工智能工厂.pdf

1、Wade Vinson,NvidiaKourosh Nemati,NvidiaUsing MGX racks and digital twins to convert any data center into an AI FactoryUsing MGX racks and digital twins to convert any data center into an AI FactoryWade Vinson-NVIDIAKourosh Nemati-NVIDIADATA CENTER FACILITY(DCF)MGX rack contributed to OCP 2024Many“ge

2、neric”but important parts needed to build AI Factory modelgeometry of the SuperPODMetadataSimReady assetOUSDSimulation aspects use case Many parts to Digital Twin,but it starts with modelNeed some experience in real life before you build the 3D ModelThe 3D Model needs EVERYTHING!No really.EVERYTHING

3、!The 3D Model and the as-built stay in sync to realize the true valueAll the parts are available to iterate rapidlyFull white space,agnostic to IT payload,with installation clearancesComplete Family of DGX Data Center Factory Buildable 3D ModulesOpen USDStand-Alone Colo or Edge DGX SuperPODNCP Scala

4、ble Module GPU plus local networkNCP Scalable GPU Module x 16NCP Scalable GPU Module x 24NCP Scalable Module Network DGX DLC and 54VNCP Scalable Module Support DGX for rPDUsSimulate CDU and POD CoolingSimulate Cluster/IT ExecutionSimulate Electrical Power BehaviourSimulateData CenterFacilities Infra

5、structureDesignOperations Digital TwinOperateTimelinePre-DesignFacility Location and site selection Physics&Workload Aware RoadmapDESIGN&SIMULATEBUILDVERIFY&OPTIMIZEDESIGNOPERATIONSDigital TwinReal Data CenterREAL TIME BMS&IT DATAMEP&IT OPERATIONSDesign AI AgentOperationsAI AgentOpen USDDIGITAL THRE

6、ADAccelerates AI factory design,build and operationsReal-time Physics Simulation&Flow Network AnalysisPhysical DataPerformance CurvesSimReady USDPhysics-Embedded AssetIntegrate in OmniverseBuild Digital TwinFlow Network AnalysisSystem-Level SimulationFuture Sc