光互連：通往 AI 開放基礎設施的途徑.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

7、agonflies104 to 205nodesMemory Pools1PCIe3Composable InfrastructureFlattened Butterflies and Dragonflies49 to 8600nodesFlattened Butterfly2Dragonfly3High Performance EthernetDragonflies333 to 60,100 nodesDragonfly A3Dragonfly B5Architectural Parameters Architectural ParameterResponseDescriptionAppli

8、cation requirement or technical capabilityOur passive optical interconnect leverages current and future optical transceiver solutions in an innovative way,delivering power and cost savingsIdentify if your assessment is an application requirement for photonics to be useful or if it is a technical cap

9、ability of your technologyUse caseAI/ML acceleration,NVMe operational flexibility,Composable Infrastructure,high-performance Ethernet,HPCIdentify the use case for your application/technology.If multiple use cases exist with significantly different parameters please describe as appropriateLinks/HostC

10、urrent topologies use 12 to 32 links per endpointDefine/identify the number of links per host ASIC or host moduleHost interface(power)Derived from chosen optical solutionDefine the electrical interface to the host and electrical interface power requirements(pJ/bit)Un-retimed/retimed?Derived from cho

11、sen optical solutionDefine if the solution is retimed or un-retimed and include CDR function if retimedFEC requirementsLink-level retransmission protects from corrected and uncorrected errorsDefine is FEC needed to support optical transmission and FEC typeBlock DiagramSee previous slidesPlease inclu

12、de a block diagram in your presentation,to support the architecture described and the parameters listed belowNumerical ParametersParameterUnitsMinMaxCommentsData rate per laneGbps10No inherent limitThe optical interconnect is flexible#lanes/link#1No inherent limitThe optical interconnect is flexible

13、Aggregate BW/linkGbps10No inherent limitThe optical interconnect is flexibleLinear BW densityGbps/mmN/AN/ADerived from chosen optical solutionAreal BW densityGbps/mm2N/AN/ADerived from chosen optical solutionEnergy efficiency single direction linkpJ/bitN/AN/ADerived from chosen optical solutionEnerg

14、y efficiency with external light sourcepJ/bitN/AN/ADerived from chosen optical solutionTotal Energy EfficiencypJ/bit20%to 60%system savingsAt scale,far fewer transceivers are requiredReachmeters1 meterNo inherent limitDerived from chosen optical solutionLatency(no FEC)nanosecN/AN/ADerived from chose

15、n optical solutionBER(no FEC)N/AN/ADerived from chosen optical solutionLatency(post FEC)nanosecN/AN/ADerived from chosen optical solutionBER(Post FEC)N/AN/ADerived from chosen optical solutionOptical channelMM,SM,PMAnyAnyDerived from chosen optical solutionOptical connectorsYes/NoYesCan use MPO,MXC

16、or any other high-density connectorMax Operating Temp CN/AN/ADerived from chosen optical solutionReliability/linkFITSN/AN/ADerived from chosen optical solutionLiquid cooling N/AN/ADerived from chosen optical solutionCost(initial volumes)$/Gbps20%to 60%system savingsSavings driven by use of lower-lan

17、e speed optics and reducing over transceiver countCost(large volumes)$/Gbps20%to 60%system savingsSavings driven by use of lower-lane speed optics and by reducing over transceiver countCommercial GA timingCYNowProven in production environments for almost 2 yearsImageLearn more about the Cerio open systems platformmattcerio.ioMatt Williams,CTO at Rockport Networks now CerioOCP Global Summit|October 18,2023|San Jose,CASYM-End