面向人工智能规模化发展的光学机遇.pdf

1、David PiehlerDell Technologieswith post-presentation editsOptical opportunities for theAI scale-up fabricOptical opportunities for theAI fabric scale-upDavid PiehlerDell Technologieswith post-presentation editsAgendaThought experiment:Scale-up the scale-up fabric and see what breaksModern copper Tra

2、nsmission media bandwidth densityShuffle and implications for DACs,AOCs,ACCs,and AECsModern opticsOptical disaggregation448GScale-up(exemplary):NVL722200 Gb/s(4 diff.pairs)NVL725,184 diff.pairs=5,184 cables per rack18 switches(144 200 Gb/s ports per switch)72 GPUs(36 200 Gb/s ports per GPU)Every GPU

3、 is connected to another GPU through the switching layerEach switch is connected to all GPUsEach GPU is connected to all switchesThought experiment:NVL72(2)2200 Gb/s(4 diff.pairs)20,736 diff.pairs=20,736 cables per rack18 switches(288 200 Gb/s ports per switch)144 GPUs(72 200 Gb/s ports per GPU;14.4

4、T)Every GPU is connected to another GPU through the switching layerEach switch is connected to all GPUsEach GPU is connected to all switches1836Double number of GPUsDouble radix of each GPUDouble number of switchesDouble radix of each switch_Quadruple number of connectionsThought experiment:512 GPU

5、clusterand shifting from 9vidia math to base2 math131,072 diff.pairs=131,072 cables128 switches(512 200 Gb/s ports per switch;102.4T)512 GPUs(128 200 Gb/s ports per GPU;25.4T)32 GPU/rack 16 switches/rackThought experiment:512 GPU cluster optical fiber!and shifting from 9vidia math to base2 math+200G

6、BASE-DR optics131,072 fibers 128 switches(512 200 Gb/s ports per switch;102.4T)512 GPUs(128 200 Gb/s ports per GPU;25.4T)32 GPU/rack 16 switches/rack8,192 fibers/GPU rack16,384 fibers/switch rackThought experiment:512 GPU cluster optical fiber!and shifting from 9vidia math to base2 math+200GBASE-DR