Comment by somnial

throughput scales superlinearly with number of GPUs when networked well and deployed with wideEP, so 1x won't compare.

also it would be interesting to figure from the DSpark paper whether their numbers are consistent with the GPUs still being H800s, since they never actually say...