Lower wattage allows more GPUs per system. For distributed training, that’s allows 8 GPUs per system which would otherwise be very difficult with retail GPUs.

And GPU-GPU interconnects are important for this type of training so putting as many GPUs as possible close together is necessary.