全新 NVIDIA A100 80GB GPU 加速超大规模 HPC 和 AI 应用.pdf

1、NVIDIA全新NVIDIAA10080GBGPU加速超大规模HPC和AI应用马四腾NVIDIA SA Team#page#AGENDA大规模HPC与AI计算面临的要求与挑战全新NVIDIAA10080GBGPU及DGXA100640GB系统NVIIDIA端到端计算平台#page#AGENDA大规模HPC与AI计算面临的要求与挑战全新NVIDIAA10080GBGPU及DGXA100640GB系统NVIDIIA端到端计算平台#page#25 YEARS OF ACCELERATED COMPUTINGDEVELOPERS+DEVELOPMENTPNINSTALLEDPERFORMANCE+BA

2、SE+ACCELERATIONCOMPUTE8CUDAEVERWHERE1NETWORKINGFull StackData Center ScaleOne ArchitectureX-factor Speed-up#page#CHALLENGES： ACCELERATING BIG AND SMALL300OxAl Interactions Per DayTring NLGalsofMedicaUSnds Adsaf pho121E-03201220132014SHOZ201720182020201620193000X Higher Compute Required to TrainEvery

3、 AI Powered Interaction NeedsLargest Models Since VoltaVaryyiing Amount of Compute#page#TODAYS HYPERCONVERGED DATA CENTERImpossible to Optimally Design Server Mix for Unpredictable Demand#page#page#AGENDA大规模HPC与AI计算面临的要求与挑战全新NVIDIAA10080GBGPU及DGXA100640GB系统NVIDIIA端到端计算平台#page#ANNOUNCINGNVIDIA A100 8

4、0GBSupercharging The Worlds HighestPerforming AI Supercomputing GPU0680GB HBMze2TB/s+Forlargest datasetsandmodelsto feed theworlds fastest GPU3rd Gen Tensor CoreMulti-instance GPU3rd Gen NVLink#page#NVIDIA A100PERFORMANCEGreatest Generational Leap- 20X Volta06PeakVsvolta31220XTF32 TRAININGTFLOPS20X1

5、.248TOPSINT8 INFERENCE19.5TFLOPS2.5xFP64HPCMULTIINSTANCE GPU7X GPUS#page#NVIDIA A100MULTI-INSTANCE GPUGreatest Generational Leap-20X Volta06Q司nd9CPUGPUGPUGPU MemGPUMeGPUMemGUMsmGPUMemGUMGPUMamUp To7hardware isolated GPU Instances perA100Simultaneous Workload Execution With Guaranteed QualityOf Servi

6、ceRight-sizedGPU#page#page#HGX AIRAPDSARAOMNIVERSEHPCSUPERCOMPUTINGPLATFORMCUDA-XPurpose-Built for the Convergence ofSimulation，Data Analytics &AlCUDAScalable from Single Server to ExascaleMAGNUMIOFull Power Of NVIDIA GPUs，NVLink，Mellanox InfiniBandFully Optimized NVIDIA Software Stack AcceleratesOv