There is no magic, if the data you compute as atomic chunk don't fit in cache then memory bandwidth R/W limit kicks in and architecture does not matter. On cont

There is no magic, if the data you compute as atomic chunk don't fit in cache then memory bandwidth R/W limit kicks in and architecture does not matter. On contrary - having multi gpu setup of same price and same memory size with even slower memories may give you effectively much higher bandwidth but at the cost of power consumption.