面向大规模人工智能的互连与连接：塑造开放式人工智能系统的互连与存储架构的未来.pdf

1、Siamak Tavallaei,Samsung SemiconductorBrian Hirano,XCENASamir Rajadnya,MicrosoftManoj Wadekar,MetaDurgesh SrivastavaInterconnects and Connectivity for At-scale AIA Panel DiscussionSERVER:COMPOSABLE MEMORY SYSTEMS(CMS)As data-intensive applications continue to scale,data-intensive workloads demanda f

2、resh look.In response,the computer industry is working on various memoryfabric-oriented proposals such as CXL,UALink,NVLink,and SUE for efficient data-movement in scale-up systems between xPUs,memory,and other end-devices.In this presentation,the panelists explore the latest trends in high-speedinte

3、rconnect starting from an internal,d2d interconnect such as UCIe toward andexternal interconnect such as photonics for at-scale AI/ML systems.They willdiscuss technical and business aspects of the trade-offs they are making for scale-up and scale-out systems.As part of the OCP/CMS effort,the panelis

4、ts will present their insights into howcomposable systems may help increase system efficiency to drive added value viadisaggregated compute,memory,and storage elements.AbstractBrian Hirano,XCENAAcademic Research and AI FabricsTRACK NAMEWhy speak with Academic Researchers?What academic research is go

5、ing on in“AI Fabrics”What ways to start engaging?What work could be done with academic researchers?AgendaPrompt:I would like to understand how next generation data center systems can use CPUs,GPUs,DPUs,CXL,PCIe,UltraEthernet,UALink,NVLink,memory and storage,so they run AI applications efficiently.Co

6、uld you draw me a diagram to show me how this can be done?Shah,A.,Chidambaram,V.,Cowan,M.,Maleki,S.,et al.,TACCL:Guiding Collective Algorithm Synthesis Using Communication Sketches,NSDI 2023.https:/research.google/blog/alpa-automated-model-parallel-deep-learning/Xinjun Yang,et,al.Unlocking the Poten