面向 GPU 加速的高性能计算和人工智能_机器学习的架构解决方案模式.pdf

1、 2025,Amazon Web Services,Inc.or its affiliates.All rights reserved.2025,Amazon Web Services,Inc.or its affiliates.All rights reserved.2025,Amazon Web Services,Inc.or its affiliates.All rights reserved.Avi KulkarniSr.WW Specialist,Accelerated ComputeWWSO Advanced Compute(he/him)Satheesh MaheswaranSr

2、.WW Specialist SA,HPCWWSO Advanced Compute(he/him)Architecting Solution Patterns for GPU-accelerated HPC and AI/MLCMP 201 2025,Amazon Web Services,Inc.or its affiliates.All rights reserved.AgendaNeed for Accelerated ComputeAccelerated Compute PortfolioHow to access provisioning mechanismsHow to crea

3、te your own infrastructure that scales elastically 2025,Amazon Web Services,Inc.or its affiliates.All rights reserved.Need for accelerated computingComplexity of workloads is driving a need for higher performanceAmount of compute acceleration needed is driving a need for higher efficiencyVarious wor

4、kloads are driving a need for multiple compute options 2025,Amazon Web Services,Inc.or its affiliates.All rights reserved.20192020202120232024/2025G4T4G5A10GG5gT4GP4A100P5H100P5enH200P6B200P6eGB200G6eL40SG6L4Broad and deep accelerated computing portfolioP6B300 2025,Amazon Web Services,Inc.or its aff

5、iliates.All rights reserved.0100020003000400050006000700005001000150020002500Instance Network bandwidth,GbpsTotal GPU Memory/Instance,GBAmazon EC2 Instance featuring NVIDIA GPUsP6,B200P5en,H200P5,H100G6e,L40SP6,B300P4de,A100P6e-GB200 not shownSize corresponds to FP16 TFLOPP4d,A100G6,G5 2025,Amazon W

6、eb Services,Inc.or its affiliates.All rights reserved.Workload DomainWorkloadRecommended Instance Other InstanceMachine Learning Large Model TrainingP5en,P6-B200,P6-B300P6e-GB200Large Model Inference/Batch inferenceP5,P5enG6e,P4Small model Inference G6,G6eG5,P4High Performance ComputingEngineering S