The memory speed would well be right if the bus is true. A 600Mhz core simply doesn't need as much bandwidth as a 800Mhz core. Having higher voltage and higher clocked memory, with huge bandwidth thats not used is a waste of power.
Say you could have 4Ghz mem at normal voltage and 550Mhz core and a huge excess of bandwidth, or drop mem speeds 20% or so, drop voltage, drop power draw of the memory by 35-40% and up core speed to 600Mhz, which would give better performance.
The problem will come in, they won't use cheap memory, but will they let you overvolt it to "normal" levels to overclock to normal memory speeds assuming you can clock the gpu to match. Thats of course assuming dropped memory speeds. At idle my power usage at the plug drops 15W or so on a 6950 going from 1250ish to 900ish. Thats without a voltage drop so maybe they'll get enough power saving without using less voltage.
The thing of it is, and the thing that winds me up about 5970 benching against the 480 then 580gtx, is its DESIGNED TO BE OVERCLOCKED. You'd be a moron not to, the 300W limit and the lowered clocks were to hit some random targets no one buying one gave a crap about, used at intended speed it was faster, most people clocked up to, or close to 5870 speeds, if you didn't you were wasting money.
Same is true about the 6990 and the same will be true about the 590gtx, it might well be slower at stock settings than a 6990 with the same powerdraw, anyone who thought otherwise when AMD provide more performance/w, was delusional.
The question is at full pelt, is it quieter, faster, or better value, which in this case would mean cheaper than 2x580gtx, and/or faster and more expensive than the 6990 in roughly the same ratio.
If its 10% faster either at stock or max realistic clocks/noise, but it costs 30% more, rubbish, if its 15% faster and costs 10% more, or Nvidia turn the screws and decide to make it quite a bit cheaper than 2x580gtx, it could be a great card.
One of the issues with the 6990 is comparing its OC'd bios with increased powertune "overclock" potential, yet not really increasing the clocks, and the memory should be clocked as [H] clearly show its bandwidth limited.
The ultimate test is how fast both are at their absolute limit, stock will be interesting but ultimately irrelevant.