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.
Been investigating these multichannel AMD based platforms last year and seem like none of them can in real scenarios utilize anywhere close to their theoretical bandwidth.