光互连：通往 AI 开放基础设施的途径.pdf

1、OCP Global Summit October 18,2023|San Jose,CASYM Title SlidePresented by Matthew Williams,CTORockport Networks Now CerioOptical Interconnect Pathways to an Open Infrastructure for AISYM-ContentCentralized Switching ChallengesIncreasing price/performance as speed growsCost of optics growing as a perc

2、entage of total systemLarge number of transceivers in each pathNon-linear transceiver cost and power at scaleCommodity MSA Optical Module Deployment8 Links of 1 Lane EachQSFP-DDs in EndpointsWith Distributed SwitchesQSFP-DDin Centralized SwitchQSFP-DDin Centralized SwitchSwitchSwitchQSFP-DDs in Endp

3、ointsWith Distributed Switches1 Link With 8 LanesTraditional Centralized SwitchingDistributed SwitchingDirect interconnectivity between endpoints 12,16,24,32 links per nodeZero power,pre-wired patch panelEliminates complexity from typical cabling Uses commodity MPO passive optical cablesTremendous f

4、lexibility and scalabilitySHFL-Passive Optical InterconnectTo NodesTo Other Racks1 RUTorusDragonflyFlattenedButterflyExample Topologies6D Torus SHFLLow-latency diameter 2 networkFull-mesh of nodes in each dimensionFull mesh in each rowFull mesh in each columnOptical pass-throughs for sparse topologi

5、es2D Flattened Butterfly262728293031323334353637383940414243444546474849251234567891011121314151617181920212223242D Flattened Butterfly SHFLCompute Nodes(13 per group)Compute/Memory GroupsDevice Chassis GroupsComposability and CXL Memory Pool ExampleLocal Group(17 nodes)Global TopologyMemory Cards(4

6、 cards per group)Dragonfly TopologyType 3 Memory Cards strategically distributed to ensure direct connectionsPCIe device chassis placed in other groups for broad sharingExample Use Cases and TopologiesUse CaseTopologiesMax Scale RangesDiameterCombined Composable Infrastructure and CXL Memory PoolsDr