Not sure if your math is right here. If NVL 576 consists of 8 racks, then 144 CPO switches should be divided into 18 CPO switches per rack, or given 9 switch trays per rack 2 CPO switches per tray, rather than your version of each tray having 6 CPO NVSwitch chips. Let me know where my math might be off.
Hey Jason thanks for your great analysis. Just a quick Q on your calculation: under "Rubin Ultra (Oberon): Phase 0 Scale-Out + Phase 1 Scale-Up", you point out CPO NVSwitch chip is co-packaged with 4.5 OEs, and on your screenshot you also say ELS Modules per GPU is 4.5. Could you please help elaborate why ELS Modules per GPU and # of OEs per chip are both 4.5? Are you using the ratio interchangeably for some reason? Thank you
Great article! Thanks for the work put into it
Thank you.
Not sure if your math is right here. If NVL 576 consists of 8 racks, then 144 CPO switches should be divided into 18 CPO switches per rack, or given 9 switch trays per rack 2 CPO switches per tray, rather than your version of each tray having 6 CPO NVSwitch chips. Let me know where my math might be off.
Hey Jason thanks for your great analysis. Just a quick Q on your calculation: under "Rubin Ultra (Oberon): Phase 0 Scale-Out + Phase 1 Scale-Up", you point out CPO NVSwitch chip is co-packaged with 4.5 OEs, and on your screenshot you also say ELS Modules per GPU is 4.5. Could you please help elaborate why ELS Modules per GPU and # of OEs per chip are both 4.5? Are you using the ratio interchangeably for some reason? Thank you