Personally, I would die to have a couple of 4870X2's to play with... Why is this not a priority?
Think about the fundamental nature of science... It pushes boundaries whenever they are set... Right now the boundary is in the TFLOPS and that might seem ridiculous, but the difference between 2.5-10 TFLOPS and 1.25 - 5 TFLOPS is quite extreme in terms of application. And moreover, because of the nature of your hardware's double precision support, it would provide substantially more power availability for those of us who would just love to see every tiny little thing in our work.
My own personal interests are in large scale simulations. Rendering of procedural planets, and physics intensive mesh modification. Such a simulator could be extended to suit the needs of many scientific communities. Whether double precision is necessary depends on the application. However the need for as much power as I can stuff in is there.
Especially in a sitation where say, NASA or any astronomy organization wants to put some data into a machine and produce a full scale precise simulation of the conditions and terrain on the planet. One would need to run many parallel calculations in order to create noise algorithms capable of simulating specific terrain based on the known scientific data. The processing power required would be enormous, so it would need every bit it could get.
If the cards are wired together in such a way that they are designed to function as one massive array... Why in the world would one want to start dividing that array?
Why can't every crossfire device expose its functionality using a single command queue?
What I would like to see is the ability to designate that a crossfire chain will be used as one CAL device. A setup that I would find very enticing would be to have 3 4870X2s in crossfire as one singular CAL device, and running along side it a 4870X2 used as a rendering device.
Don't sell us short here... Besides wouldn't the company have more to gain if I were to buy 4870X2's and not just 4870's?
If the idea is massively parallel processing... Well then why sell out the most massive.