在 AWS 上为 AI 工作负载构建网络 [重复].pdf

1、 2025,Amazon Web Services,Inc.or its affiliates.All rights reserved.2025,Amazon Web Services,Inc.or its affiliates.All rights reserved.N E T 4 0 4-RArchitecting networks for AI workloads on AWSRiggs Goodman IIIPrincipal SA AI SecurityAWSJohn PanglePrincipal Product Manager TechnicalAWSBrian BarrettS

2、r Principal TechnologistAWS 2025,Amazon Web Services,Inc.or its affiliates.All rights reserved.Chalk TalksDisclaimer around chalk talks at re:InventA highly interactive content formatBegin with a short presentation(THATs THIS)followed by an open-format Q&A sessionGoal:chalk talks foster technical di

3、scussions about real-world architecture challengesUp to you(the audience)to ask questionsif not,we will just walk through some architecturesThis is L400,which means we will go from L100 to L400 level contentas long as you ask itWe wont share anything that isnt public,but can follow up with you after

4、wards 2025,Amazon Web Services,Inc.or its affiliates.All rights reserved.Model InferenceAI networking to interact with the model once trainedModel TrainingAI networking for creating a model to useData TransferAI networking to get the massive amounts of data to the model for inference or trainingUse

5、the modelThree different areas to exploreUse the modelData to the model 2025,Amazon Web Services,Inc.or its affiliates.All rights reserved.What are customers looking for?With AI networking,different requirements exist compared to traditional workloads,across inference,training,and data transfer.High

6、 bandwidth:involve massive datasets and elephant flows between GPUsUltra-low latency:Slight delay in data delivery can stall training job or bad user experience with inferenceNo packet loss:Even minor packet loss can dramatically degrade performance and increase training job completion timeScalabili