AI xPU 规模化扩展的光学技术考量.pdf

1、Rang-Chen(Ryan)YuOptical Technologies for AI Scale UpOptical Technologies for AI Scale UpRang-Chen(Ryan)YuTeraHopOPTICSAI Data Center Networking InfrastructureGeneral Purpose Servers and StorageScale-Out Front-End NetworkScale-Across Scale-Up GPUs Scale Up and Scale Out Bandwidth Growth TrendScale-O

2、ut bandwidth 10 x of scale outScale up 10Tbps,scale out 800G-1.6TTodays main-stream AI Scale-Up with copper cables interconnecting GPUs within rack AI Scale Up Networking Going from Copper to OpticalIndustry progress towards optical scale-up for multi-rack SuperPodGoogle Ironwood TPU Superpod:9216 c

3、hips per podCopper within rack,optical across rack,with OCS as switching UALink:New standards for open Scale Up Supports up to 1024 xPUsWill requires optics for more than 1 rack Market Forecast for Optical Scale-UpLight Counting starts forecasting meaningful Scale-Up Optics 2026-2027+Scale-Up Optica

4、l Connectivity RequirementsBandwidth:GPU scale-up requires 10 x Scale-out higher bandwidthe.g.,for NV GB200,scale-up bandwidth per GPU 7.2Tbps,vs.Scale out NIC ConnectX-8 800Gbps ReliabilityScale-up network bounds large number of GPUs as one memory coherent domain,reliability critical for high GPU u

5、tilization rate:10 FITLatencyLow latency important to keep GPU SupertPod performance,also limit numbers of racks within single scale-up domain(fiber delay 35ns/10m)PowerHigh bandwidth optics power 40B device-hours SiPho PIC FIT rate estimate 0.03 FITDevice-HoursPLO trending to$0.1 with 1.6T TRO 2030

6、(Light Counting)Further cost reduction potential with higher degree of silicon photonics integration(e.g.,12.8T=64x200G PLO)Pluggable Optics(PLO)ASP Trend$1.1 40Gbps$0.1 1.6T(TRO)at 203010 x cost reductionOver 40 x BW increasePluggable Optics(PLO)Power Trend051015202530352016201820202022202420262028