用于规模化生产的光学器件.pdf

1、Rang-Chen(Ryan)YuTeraHopOptics for AI Scale UpOptics for AI Scale UpRang-Chen(Ryan)YuTeraHopOPTICSAI Data Center Networking NICGPUScale Up 10 x BWScale OutScale AcrossGPU SuperPodScale Up RequirementxPUScale UpGPU SuperPodScale Up SwitchBandwidth:Scale-up requires 10 x higher bandwidthe.g.,for NV GB

2、200,scale-up bandwidth per GPU 7.2Tbps,vs.Scale out NIC ConnectX-8 800Gbps Reach and connectivity typeGoing from 1 rack to multiple rackCopper cable is liming,need to go opticalReliabilityLarge number of GPUs as one memory coherent domain,reliability critical for high GPU utilization rateLatencyLow

3、latency important to optimize GPU Super Pod performanceHigh performance with low BER with low latency FECPowerHigh bandwidth optics power 40Bilion device hours deployed with PLO,PIC FIT 0.1FITDevice-HoursOptical Transceiver Based on Silicon PhotonicsSilicon Photonics Tx Eye-200G PAM4Silicon photonic

4、s based 1.6T optical transceivers with proven performance at 200G-PAM4Hybrid silicon photonics with TFLN promising to scale to 400G-PAM4TFLN Tx Eye at 400G PAM4BER for a 1.6T-2x DR4 TransceiverBER=10-10Ch.#1Ch.#2Ch.#3Ch.#41.6T-RTLR reduce power to 16 W with 3nm DSPEarly testing with commercial switc

5、h showing promising results with low BER floorOIF ECOC25 Interop Demo Silicon Photonics Supporting Linear OpticsPluggable Optics(PLO)Power Trend0510152025303520162018202020222024202620282030400G-FRO800G-FRO1.6T-FRO400G-LPO800G-LPO1.6T-TRO1.6T-LPODSP progress towards 2nm CMOS will drive 1.6T-TRO 7.5p

6、j/b1.6T-LPO may reach 5pj/bPLO trending to 10Tbps/moduleNPO(Near Package Optics):Closer to ASIC to Reduce e-channel loss“open box repair/replaceOption supporting“open socket”for healthy eco-systemCPO(Co-Package Optics):Lowest e-channel loss with optics on same