面向NVIDIA加速计算的下一代冷却技术.pdf

1、Next-Generation Cooling For NVIDIA Accelerated ComputingAli Heydari,Director,Data Center Cooling&InfrastructureAug 25,2024Era of AIApplication FrameworksThe 3 connected dynamics to give Million-X leap in computationsAIACCELERATED COMPUTINGDATA CENTERS The 3 connected dynamics AI,Accelerated Computin

2、g and Data Centers are revolutionizing the way science is done.The million-X leap can solve problems previously impossible,like in the areas of computational fluid dynamics,data science,climate,imaging,computational biology,computational lithography,quantum physics and many more.Human BrainFrontierM

3、aximum performance(Pflop/s)10001206Power Usage(kW)0.01522,786Energy Efficiency(Gflops/w)66,666,66652.93101001000PflopsFP64HumanbrainGenerative AI models to drive the next industrial revolutionLarge AI factories required for:Hosting multiple GPUs with high rack densityProviding higher throughput,low

4、latency and running larger AI models in real-timeDriving growth in tokens/sec&revenueComplexity&scale of modern AI models require significant computational power.Multi GPU clusters in AI factories are the future for producing AI tools.Liquid Cooling to enable NVIDIAs next-gen chips like Blackwell pu

5、sh boundaries in both training and inferenceBuilding Bigger AI factoriesSelene 20214,480 A100 GPUs3 EF AI Compute112 TB/s BWEOS 202310,752 H100 GPUs43 EF AI Compute1100 TB/s BWNext AI Factory32,000 GPUs645 EF AI Compute58,000 TB/s BWEF-ExaFLOPsBW-BandwidthFactories for Generative AI:Data CentersHybr

6、id Data Centers(Air+Liquid Cooling)Cold aisle containment with supply of air through perf tiles in the cold aisles,return of the air to the CRAHs in the hot aisle Liquid supply(PG25)to the rack manifolds using QDs to servers Flow distribution through the data center using row manifolds and headers f