面向人工智能网络的横向扩展以太网架构.pdf

1、Arihant Jain,AristaBhaskar Chinni,BroadcomScale-out Ethernet Architectures for AI NetworksEmbargoed till Oct 13th,2025,9 am Pacific TimeScale-out Ethernet Architectures for AI NetworksArihant Jain,AristaBhaskar Chinni,BroadcomOCP Special Focus:Artificial Intelligence(AI)Introduction Why Scale-out AI

2、 Ethernet Networking ArchitecturesSwitch Scheduled vs Endpoint ScheduledRail&Plane ArchitecturesSAICall To ActionAgendaWhy Scale-Out AI Networking?Evolution from an era of information to an era of intelligence is creating unprecedented demand for AI InfrastructureAccelerating the evolution of AI net

3、working to deliverunparalleled scale(Mn+GPUs)efficiencyflexibility(overcoming Hetrogenties-GPU/xPU)Traditional Clusters vs AI Clusters AI Networking Big PictureEndpoint Scheduled/Non Scheduled Fabric(NSF)Low Latency,higher switch capacitySimplified CablingSwitch Scheduled Fabric/Disaggregated Switch

4、 Fabric(DSF)NIC Agnostic Deep Buffer Similar topology choices possible with both designs:Top of Rack(ToR)/End of Row(EoR)Rail Based/Planar Single Stage,Two Stage,Three stage Clos et.al.Design ChoicesEndpoint Scheduled/Non Scheduled FabricLeafLeafLeafLeafSpineSpineSpine.Standard Ethernet network conn

5、ectivity Leaf/Spine/Super-Spine:switching,forwarding,queuing,scheduling High bandwidth and radix;flat topology Adaptive routing and support for e2e congestion management Endpoints implement end-to-end congestion management and spraying/load balancingSuper SpineSuper SpineSwitch Scheduled/Disaggregat

6、ed Switch FabricLeafLeafLeafLeafSpineSpineSpine.Standard Ethernet network connectivityNIC Agnostic SolutionLeaf:switching,forwarding,queuing,schedulingSpine/Super Spine:forwarding at low powerLossless delivery from ingress to egressCell spraying ensures no con