6590 - Balancing Compute Memory and Networking for AI (Broadcom).pdf

《6590 - Balancing Compute Memory and Networking for AI (Broadcom).pdf》由会员分享，可在线阅读，更多相关《6590 - Balancing Compute Memory and Networking for AI (Broadcom).pdf（11页珍藏版）》请在三个皮匠报告上搜索。

1、FROM IDEAS TO IMPACT!SVP&GM,Core Switching GroupBroadcomRam VelagaCopyright 2025 Broadcom.All Rights Reserved.The term“Broadcom”refers to Broadcom Inc.and/or its subsidiaries.ComputeMemoryNetworkingBalancing compute,memory&networking for optimal performanceScale-up for AICompute,Memory&Networking fo

2、r AIXPUHBMCopyright 2025 Broadcom.All Rights Reserved.The term“Broadcom”refers to Broadcom Inc.and/or its subsidiaries.AI Scale-up and Scale-out NetworkingSpineLeafScale-upScale-outCopyright 2025 Broadcom.All Rights Reserved.The term“Broadcom”refers to Broadcom Inc.and/or its subsidiaries.XPUHBMHBMH

3、BMHBMXPUHBMHBMHBMHBMXPU Scale-up:High Bandwidth Memory Sharing Across XPUs4 x HBM3E(9.6Tbps)38.4Tbps8 x HBM4(12.8Tbps)102.4TbpsCopyright 2025 Broadcom.All Rights Reserved.The term“Broadcom”refers to Broadcom Inc.and/or its subsidiaries.XPU Scale-up RequirementsHigh bandwidthEfficient data transferRe

4、liable connectivity123XPUHBMHBMHBMHBMCopyright 2025 Broadcom.All Rights Reserved.The term“Broadcom”refers to Broadcom Inc.and/or its subsidiaries.Copyright 2025 Broadcom.All Rights Reserved.The term“Broadcom”refers to Broadcom Inc.and/or its subsidiaries.Addressing Scale-up NetworkingVery high bandw

5、idth per XPUReliable and power efficient interconnect(copper or fiber)Directly connecting 512+XPUsHigh bandwidth Ethernet switches:50Tbps 100Tbps 200Tbps Advanced semiconductor manufacturing 5nm 3nm 2nm+Increasing SerDes lane speeds:100G 200G 400GNot constrained to one rackCopper,CPC,lower power plu

6、ggables&CPOCopyright 2025 Broadcom.All Rights Reserved.The term“Broadcom”refers to Broadcom Inc.and/or its subsidiaries.Addressing Scale-up Networking(contd)Low-latencyLink level retryOptimized headersPFC or credit-basedflow controlEfficiencyReliabilityCopyrig