争论现状：可插拔线性光学器件与共封装光学器件.pdf

1、State of the Debate:Pluggable and Co-Packaged Optics Vlad Kozlov,LightCountingDemand for optics in AI ClustersSource:Optics for Cloud Data Centers July 2025()Transceivers and GPUs sold by NvidiaSource:Optics for Cloud Data Centers July 2025()Complex correlation between the numbers of GPUs and Transc

2、eivers:NICs,switch SerDes and more.-2.00 4.00 6.00 8.00 10.00 12.00202220232024202520262027202820292030Units(M)400G800G1.6T3.2T-1.00 2.00 3.00 4.00 5.00 6.00202220232024202520262027202820292030Units(M)HopperBlackwellBlackwell UltraRubinRubin UltraFeynmanSwitches for Scale-Out and Scale-Up-200.00 400

3、.00 600.00 800.00 1,000.00 1,200.00 1,400.00 1,600.0020232024202520262027202820292030SerDes lanes(M)50G100G200G400G-200.00 400.00 600.00 800.00 1,000.00 1,200.00 1,400.00 1,600.00 1,800.0020232024202520262027202820292030SerDes lanes(M)100G200G400GSource:Optics for Cloud Data Centers July 2025()Scale

4、-Out:Ethernet and InfiniBandScale-Out:NVLink and otherScale-Up First,Scale-Out LaterSource:Nvidia GTC 202540 x increase in performance!Mostly because of the larger number of GPUs in the scale-up domain:From 8GPUs in NVL8 to 72 GPUs(or 144 GPU dies)in NVL72Google and Nvidia architecturesBoth systems

5、deliver similar performance,but Googles scale up domain is 9,216 TPUs vs.72 GPUs of Nvidia,but Googles ICI between TPUs connectivity is not as“good”as NVLink(no memory sharing).Nvidias system requires almost 2.7x more transceivers(without using optics in scale-up networks).Googles Ironwood Pod compr

6、ises 9,216 TPUs with an aggregate 42.5 exaflops of FP8 performance.It requires 1.5 of 800G transceiver per TPUNvidia would need 118 of its Blackwell Ultra NVL72 racks,which would have 8,496 GPUs in total to reach 42.5 exaflops of FP8 performance.It requires 4 of 800G transceivers