If you are interested in playing, any card 8xxx series or above works. (Nvidia) For instance, this Macbook has 9400M and I was able to download the CUDA stuff and run the sample programs with no trouble. RE the double precision and # of shaders (or performance), it really depends on what your code uses and how fast you need/want it. Easiest to get something running and then see where the bottleneck is, and how much it costs to fix.
Mostly my GPU usage is for folding (folding.stanford.edu), I like the new boards because they run cooler. Just ordered a Gt240 for about $65 after rebate. An advantage is that the 240 doesn't need the extra power cable. I think it will fold proteins as fast as about ten Core2 2Ghz CPUs. The best card I have so far is a 250, it equals about 20 of those Core2's. Last summer I picked up some 9600gso cards for about $35, those have similar performance to the 240, but require the extra power plug.
I would like to do something like atlasfolding.com , but with much less $$. It looks like this new 480 is about 4X's the 295 performance for ~same cost. Sounds good to me.
PS - if you get some good GPGPU code running and need more performance, try to hook up with a .edu HPCC. Most of them are getting into CUDA and might have spare cycles. You might have to switch to linux.