面向人工智能_机器学习加速器的可扩展以太网（SUE）框架.pdf

《面向人工智能_机器学习加速器的可扩展以太网（SUE）框架.pdf》由会员分享，可在线阅读，更多相关《面向人工智能_机器学习加速器的可扩展以太网（SUE）框架.pdf（13页珍藏版）》请在三个皮匠报告上搜索。

1、Mohan Kalkunte,Vice President,Architecture,BroadcomHugh Holbrook,Chief Development Officer,Arista NetworksSUE:Scale Up EthernetEmbargo until October 13,2025SUE:Scale Up EthernetMohan Kalkunte,Vice President,Architecture,BroadcomHugh Holbrook,Chief Development Officer,Arista NetworksOCP SPECIAL FOCUS

2、:ARTIFICIAL INTELLIGENCE(AI)AI Scale-up and Scale-out NetworkingSpineLeafScale-upScale-outXPU Scale-up:High Bandwidth Memory Sharing Across XPUsXPUHBMHBMHBMHBMXPUHBMHBMHBMHBM4 x HBM3E(9.6Tbps)38.4Tbps8 x HBM4(12.8Tbps)102.4TbpsKey Requirements:High Networking Bandwidth,Efficient Data Transfer,Reliab

3、le TransportEthernet Scale-up:High Performance,Open,Existing SpecificationsXPU BXPU AXPU CSUE-TransportOther Accelerator Implementation Accelerator ImplementationEthernet PHYEthernet Data LinkEthernet HeaderFocus AreasL2/L3 transportEfficient headersError recoveryLossless networkExecute at your own

4、paceFreedom to innovate/implementEthernet for Scale-up Networking(ESUN)Transport over EthernetIntroducing SUE-TransportSUE-TransportAccelerator A protocolAccelerator B protocolXPUXPUXPUXPUA Transport over Ethernet can choose:Core-side interfacePush vs pull memory accessTransaction packing policyTran

5、sport vs hop-by-hop reliabilityOrdering modelCongestion control approachTransaction sizesEncryptionLoad BalancingScheduling.SUE-Transport is one set of choicesamong these optionsSUE-Transport Focus AreasIntegration:RDMA NICs are too large to integrate many instances into an XPU package.scale-up band

6、width is commonly 8-12x scale-outBandwidthMap memory read/writes and atomics onto Ethernet packetsMemory SemanticsHundreds of XPUs Thousands of XPUsSingle-hop Multi-hopIn-order deliveryTopologyExisting solutions(UEC,RDMA,TCP)not optimized for scale-upThey generally solve a harder