Originally Posted by MUYA
Could be a bottleneck somewhere in the actual GPU/engine?
Memory bandwidth would be my guess. 1200-->1300 at 256 bit isn't much bandwidth increase(+12.8GB/s). However, on the 480 they used here, 924 --->1100 at 384 bit is more than twice as much bandwidth gain, about 2.5 times or so(+33GB/s). Mainly because NVs GDDR5 memory controller sucks, they are actually downclocking the slowest GDDR5 to make it work, but it leaves them with more memory overclocking headroom.