用于 ML 阵列边缘的高密度光互连.pdf

1、OCP Global Summit October 18,2023|San Jose,CAKaren Liu,Nubis CommunicationsHigh-Density Optical Interconnect for ML ClustersIntra-cluster network needs to look more like chip-to-chip I/O than networking I/OTechnology Gap for Pod-Pod within a ClusterTo external networkChip-to-chipLow power/small die

2、areaExternal networking Traditional copper/optics Single xPULocal PodFull ClusterCluster I/O needOptical reach with chip edge density Pod-pod interconnect30 mm24 dB each end)Architectural ParameterDescription(Capability)Application requirement or technical capabilityApplicationTechnical CapabilityUs

3、e caseShelf-shelf interconnect between xPU within a cluster.High-density Linear I/OLinks/Host32 144 lanes/chip 2x2 to 5x5 chips per local pod“substrate”2-3 links per pod,depends on topology100 Gbps/laneAny aggregation of lanes per linkHost interface(power)32 Gbps,64 Gbps,100 Gbps PAM4;56 NRZ No hard

4、 power cut-off,but less is betterMR,LR SerdesUn-retimed/retimed?Unretimed preferred for latency&powerUnretimed preferred for densityFEC requirementsFEC required,KP4 mostly,varies by application KP4 is defaultParameterUnitsMinMax(Capability)&AssumptionsData rate per laneGbps25200#lanes/link#1632Aggre

5、gate BW/linkGbps16003200Linear BW densityGbps/mm1005001 link per outward facing edge chip edgeAreal BW densityGbps/mm231015-20 mm edge width x 30mm reticle dimensionEnergy efficiency single direction linkpJ/bitLess is better10 (4.5 typ)Energy efficiency with external light sourcepJ/bitLess is better

6、 10 (6)External laser is in 3(500)“Once optical,people will likely want more reach”Latency(no FEC)nanosec1 n“Same as copper for same reach.”BER(no FEC)1e-8includes 2 orders of magnitude desired for marginLatency(post FEC)nanoseclatency with