以太网不断发展借助超级以太网联盟为人工智能的未来提供动力.pdf

《以太网不断发展借助超级以太网联盟为人工智能的未来提供动力.pdf》由会员分享，可在线阅读，更多相关《以太网不断发展借助超级以太网联盟为人工智能的未来提供动力.pdf（21页珍藏版）》请在三个皮匠报告上搜索。

1、1|2024 SNIA.All Rights Reserved.AMD Official Use Only-GeneralEthernet Evolved:AIs Future with the Ultra Ethernet Consortium(UEC)Presented by:J Metz,Ph.DChair,UECJ Metz,Ph.D,2|2024 SNIA.All Rights Reserved.AMD Official Use Only-General“Ultra Ethernet?”-Agenda Why“Ultra Ethernet?”Who is Ultra Ethernet

2、 Consortium(UEC)?Philosophy of UEC Overview of projects Where to get more information23|2024 SNIA.All Rights Reserved.AMD Official Use Only-GeneralAI for Networking,or Networking for AI?Many articles/blogs have talked about how AI can change the networking infrastructure but what network infrastruct

3、ure do you need to have enough AI to change the networking infrastructure?Is it more than just superfast speeds and feeds?Massive data sets,parallel processing requirements Where does the data need to be,and when?4|2024 SNIA.All Rights Reserved.AMD Official Use Only-GeneralThe AI Monster AI workload

4、s need Ever-increasing Memory Bandwidth Ever-increasing Memory Capacity(Near)Instantaneous Data Access(Exabytes)Intermittent data surges”Straggler”data(tail latency)significantly impacts completion time Extended operation duration(hours,days)5|2024 SNIA.All Rights Reserved.AMD Official Use Only-Gene

5、ralWhich Network?UEC=Scale-Out6|2024 SNIA.All Rights Reserved.AMD Official Use Only-GeneralAI Problems To Solve Memory Bandwidth vs.Capacity vs.Latency Computation-Bound Workloads E.g.,Summarization:processes large input context simultaneously(parallel)Higher weights reusability Memory-Bound Workloa

6、ds E.g.,Generation:produces single word at a time(iteration)Lower weight reusability Recommendation Workloads spend almost 60%of time in Network I/O*Meta,OCP 2022 Global Summit7|2024 SNIA.All Rights Reserved.AMD Official Use Only-GeneralThe Goodput Dilemma Logarithmic expansion of the relationships