针对您的 AI_ML 工作负载提供 AWS 计算选项的战略指南.pdf

1、 2025,Amazon Web Services,Inc.or its affiliates.All rights reserved.2025,Amazon Web Services,Inc.or its affiliates.All rights reserved.A I M 3 6 7A strategic guide to AWS compute options for your AI/ML workloadAnoop SahaSr GTM Specialist,Gen AIAWSMichele MonclovaPrincipal Product Manager,AI Platform

2、sAWS 2025,Amazon Web Services,Inc.or its affiliates.All rights reserved.AgendaKey considerations for your training and inference workloadsDiving deeper into GPU architectureHow to select the right instance type for your workload How to get the right instances at optimal cost 2025,Amazon Web Services

3、,Inc.or its affiliates.All rights reserved.0How many times did Matt Garman mention“Generative AI”in his Re:Invent 25 keynote?144What is the maximum number of trainium chips in a single ultraserver?96How many times did he mention“AI”?10How many EC2 accelerated compute instance families did AWS announ

4、ce in last 2 years?If an AI model is answering these questions after analyzing the video,which instance family should they use?2025,Amazon Web Services,Inc.or its affiliates.All rights reserved.Fundamental architectures for different workloadsTrainingPlacement groupAvailability ZoneAmazon FSx for Lu

5、streTightly coupled,communication heavy&inter-node latency sensitiveOptimization:Time to trainEKS node groupAvailability Zone 1Availability Zone 2InferenceLoosely coupled,fast scaling in/out&query latency sensitiveOptimization:Latency,Throughput,ScalabilityCustomizationHeterogenous cluster,communica

6、tion heavy&inter-node latency sensitiveOptimization:Iteration timeAvailability ZoneGPUGPUCPUAmazon FSx for Lustre 2025,Amazon Web Services,Inc.or its affiliates.All rights reserved.Accelerated compute instances on AWSG4NVIDIA T42019P4NVIDIA A1002020G5gNVIDIA T4GG5NVIDIA A10G2021P5NVIDIA H1002023G6eN