以太网扩展概述.pdf

1、ScaleScale-up Ethernet up Ethernet OverviewOverviewBroadcomNETWORKINGScale-up Ethernet OverviewRam VelagaRam VelagaSVP/GM,BroadcomCompute,Memory&Networking for AIComputeComputeMemoryMemoryNetworkingNetworkingBalancing compute,memory&networking for optimal performanceBalancing compute,memory&networki

2、ng for optimal performanceScaleScale-up for AIup for AIXPUHBMAI Scale-up and Scale-out NetworkingSpineLeafScale-upScale-outXPU Scale-up:High Bandwidth Memory Sharing Across XPUsXPUXPUHBMHBMHBMHBMXPUXPUHBMHBMHBMHBM4 x HBM3E(9.6Tbps)38.4Tbps8 x HBM4(12.8Tbps)102.4Tbps6High bandwidthHigh bandwidthEffic

3、ient data transferEfficient data transferReliable connectivityReliable connectivity1 12 23 3XPUXPUHBMHBMHBMHBM6|XPU Scale-up RequirementsAddressing Scale-up NetworkingVery high bandwidth per XPUReliable and power efficient interconnect(copper or fiber)Directly connecting 512+XPUsHigh bandwidth Ether

4、net switches:50Tbps 100Tbps 200TbpsAdvanced semiconductor manufacturing 5nm 3nm 2nm+Increasing SerDes lane speeds:100G 200G 400GNot constrained to one rackCopper,CPC,lower power pluggables&CPOAddressing Scale-up Networking(cont.)Low-latencyLink level retryOptimized headersPFC or credit-basedflow con

5、trolEfficiencyReliabilityAdaptable Optimized Headers:Low-Overhead ForwardingIncreased network efficiencyFully Ethernet CompliantConfigurable for different applicationsReduced packet overheadPayloadStandard IPv4/UDP Header46BOptimized Ethernet Header10BPayloadThe Adaptable Optimized Header is compati

6、ble with any standard Ethernet switchThe Adaptable Optimized Header is compatible with any standard Ethernet switchLossless Fabric:Link Layer Retry(LLR)ApplicationTransportIPLink LayerApplicationTransportIPLink Layerx(1)FEC Error(2)Request retry(3)RetransmitNo loss detectedSender