基于NVIDIA GPU的新一代液冷AI数据中心创新解决方案.pdf

1、Innovative Solutions for New Generation Liquid-Cooling AI Data Centers Based on NVIDIA GPUsJie HuPE,R&D Department,AivresAI DCDevelopment Trend12Challenges in AI DC InfrastructureThe computing power and power consumption in the data center space continues to increase.Driven by the development of new

2、 AIGC business model,the construction of AI DCs faces many challenges.Such challenges include the needs of large models,an increase in IT density,and rising energy consumption/cost.The traditional IT and data center infrastructure has a long construction period.Dilemma:new data center outdated as so

3、on as it is built.3-5 years3-5 years3-5 years10-15 yearsIT Life CycleDC Life cycleConstruction Period:3 YearsSource:Based on actual vendor data3Innovative Solutions24AI Data Center ArchitectureCooling MachineCooling Source and Water Treatment DeviceOpen towerIntegrated Cooling SourceLiquid Cooling C

4、abinetCDULiquid-AirLiquid-LiquidLiquid Cooling componentsSecondary Cooling LoopAir-Cooled Air ConditionerMicromodule(Closed Loop)5Row-Level Computing Power Modules:Large-Scale and Efficient DeploymentPOD Level DeploymentSingle-row/double-row POD design,single POD uses 16 cabinets for deployment.CDU

5、uses 2N/N+1 deployment for high reliability.Wide Geographical DeploymentUses high performance heat exchange megawatt-class CDU.CDU extreme performance mode solves the problem of having a high water temperature in tropical areas.Safety and ReliabilityUpper pipe routing,two-layer design provides conve

6、nient maintenance.Uses water collection tray and leakage monitoring system.Supports pipeline leakage monitoring.Supports removal of the cabinet system when leakage occurs thus improving liquid cooling safety.6Prefabricated Computing Box Improves Deployment EfficiencyHighly IntegratedThe 20-foot cont