Originally Posted by badass 
Before anyone lays into me, if the cards are designed right, the copper paths between the GPU's could be short and noise free enough to easily run as fast as the memory interfaces themselves. There also wouldn't need to be much of a latency hit for accessing memory connected to the other GPU.
But there's no point, because the other GPU would be using it first
Unless, of course, you have one GPU idle while the other is doing all the work, that'd be efficient
I love the OP comparison of current card vs yet to be released card. We love them here
Now get the nvidia yet to be released card vs ATI yet to be released card.. that'd be a nice comparison.