网络：面向 ML AI 的超高密度、超低功耗 16x112Gbps 线性驱动硅光子学.pdf

《网络：面向 ML AI 的超高密度、超低功耗 16x112Gbps 线性驱动硅光子学.pdf》由会员分享，可在线阅读，更多相关《网络：面向 ML AI 的超高密度、超低功耗 16x112Gbps 线性驱动硅光子学.pdf（13页珍藏版）》请在三个皮匠报告上搜索。

1、OCP Global Summit October 18,2023|San Jose,CAPeter Winzer,Nubis Communicationswww.nubis- Ultra-Dense and Ultra-Low Power 16x112-GbpsLinear-Drive Silicon Photonics for ML/AIOptical Package I/O For Advanced ML/AI ClustersHDI/O(High-density I/O for ML/AI)Ultra-high I/O densityTbps/mmUltra-low latency1

2、ns add-on over passive copperUltra-low power 6 pJ/bit(incl.laser,control)Intra-rack to inter-row 1 meter to 100s of metersCurrent I/OASICCurrent I/O Densities Dont Match ASIC NeedsASIC1:5fanout Fiber edge coupling limits density Fan-out increases trace length(channel loss)Needs WDM to scale:not idea

3、l for ML/AI interconnectFull-Duplex I/O DensityTransmission ReachHostI/ODie-to-die500 Gbps/mm 5 Tbps/mm2 mm substrate traceSerDes(100G,200G)500 Gbps/mm 1 Tbps/mm25-36 dB channel lossSystem I/OCopper Cable40 Gbps/mm 150 Gbps/mm1 2 mCurrent DR Optics80 Gbps/mm 250 Gbps/mm100s 1,000s mScalable Optical

4、HDI/O Needs 2D SolutionsFull-Duplex I/O DensityTransmission ReachHostI/ODie-to-die500 Gbps/mm 5 Tbps/mm2 mm substrate traceSerDes(100G,200G)500 Gbps/mm 1 Tbps/mm25-36 dB channel lossSystem I/OCopper Cable40 Gbps/mm 150 Gbps/mm1 2 mCurrent DR Optics80 Gbps/mm 250 Gbps/mm100s 1,000s m2D Optical HDI/O2

5、50 Gbps/mm multi-Tbps/mm100s 1,000s mASICCtrl2D HDI/O 2D HDI/OASIC 2D fiber array overcomes beachfront limitation Vertical fiber escape enables 2D HDI/O tiling No WDM needed Lower complexity,power,cost Higher architectural flexibility(full fan-out)ASICDataProcessingRetimerDSPSerDesSerDesSerDesOEOEOE

6、OEOEOEOEOEPower&Latency:Optics Must Look Like CopperCarefully designed opto-electronics are linear “Look like copper”to Host SerDes Eliminates Retimer DSP No additional latency relative to copper(2 x DSP adds 200 ns)Reduces I/O real estate and power consumption by 50%Copper Cable