超越规范：机架级部署 UALink 需要哪些条件.pdf

1、Chris Petersen,Fellow,Technology&Solution ArchitectureAhmad Danesh,AVP,Product ManagementAstera Labs Beyond the Spec:What it Takes to Deploy UALinkat Rack-ScaleBeyond the Spec:What it Takes to Deploy UALinkat Rack-ScaleChris Petersen,Fellow,Technology&Solution ArchitectureAhmad Danesh,AVP,Product Ma

2、nagement Astera Labs OCP SPECIAL FOCUS:ARTIFICIAL INTELLIGENCE(AI)Scale-Up Solution Requirements123456NeedSolutionFast TTMData PrivacyApp UptimeDeployableScalableTCO OptimizedInteroperabilityRASSecurityScale-Up RacksRack MgmtUALinkUltra Accelerator Link (UALink)Board of DirectorsOpen,standard,accele

3、rator-to-accelerator communicationSimple memory semantic based protocol for up to 1k XPUs Switch,Node,and POD management Data rates of Ethernet,low latency of PCIe,low power Efficient for silicon implementation and fast time-to-marketSwitchOpen collaboration to deliver purpose-built scale-up protoco

4、l for AI at rack-scaleOct 2024April 2025Coming SoonConsortium IncorporatesUALink 200G 1.0 Spec.UCIe IO Chiplet Spec.In-Network Collectives Spec.Management Spec.UALink 200G Scale-Up RackUALink ChipsUALink PlatformsUALink 200G Rack UALink-based XPUs UALink 200G Switches Ethernet Retimers UALink Chips

5、Cables/Backplane Connectors DC-SCM for Management UALink Platforms Common Rack Infrastructure UALink POD ControllerETH RetimersRack ControllerRack ControllerRack ControllerUALink XPUsIndustry is actively developing silicon and collaborating on common platform&rack infrastructureUALink SwitchesPCIe1/

6、10GbEUALink POD ControllerRack-level configScale-up orchestration for XPUs&SwitchesID&vPOD assignmentNode Mgmt Agent(NMA)Runs on BMCUALink topology discoveryGPU configurationSwitch Mgmt Agent(SMA)Runs on CPU or BMCSwitch-level configID&routing table configUALink Rack-Level Management Architecture 12