Effect of memory bandwidth on GPU performance

A Radeon 6970 ships with GDDR5 that can run at 6Ghz effective rate, but it only runs it at 5.5Ghz (though there are factory-overclocked cards that reach at least 5.7Ghz). NVidia's cards tend to run their memory at no more than 4.5Ghz effective, though they compensate by using a wider memory bus (which accounts for the odd-looking memory sizes).

Upper-mid-range cards built off smaller cores tend to have higher core clocks, but much lower memory speeds. Halving the number of compute units decreases the demand for memory bandwidth enough that it's not crippling when you use a memory controller that's half the size and can only run at 75% the speed, which saves the manufacturer a lot of money.