The GT 240 is good choice for small PSU's. Note that you might also find one of the later 9800GT "green" models that will run a bit faster with similar power requirements.

Shader clock is the most important thing, but you can probably overclock the 550 to 585 easily (factory OC cards are a good choice since they have been stress tested at the higher rates already). I went with the 512 of GDDR5 when choosing my GT 240. That was back on the AP26 app, and the faster memory did provide a small extra boost in speed with that application. Whether that is true of the current PPS sieve app is not yet clear. 512 is also plenty for the applications which can easily run on 256 without issue, so that is not a decision point.


____________
141941*2^4299438-1 is prime!

I wouldn't worry about the compute capability here. There is very little difference in 1.1 vs. 1.2 (GPUGRID is the only project I know where that matters). Here at PG, the PPS app can use any compute capability, even the oldest 1.0 cards. The old AP26 app needed only 1.1 capability. I would not expect PG would require it any time soon if ever, but some of the developers would be better at answering that question definitively. Also, 1.2 is only single-precision just like 1.1; you would need 1.3 or one of the Fermi cards for double precision (and on a side note, the Nvidia DP is much, much slower than ATI's, so projects needing it generally develop ATI apps as well).

As for the 9800GT being "much" faster than the GT240, I'd say not so much. The GT240 shader clocks are really low (1340 as I recall) compared to higher clocks for the 9800GT. If you could clock the 240 up to the 1700-1800 range, the difference--though still to the advantage of the 9800--would be much more modest. For example, the 96 shader 9600 GSO on the list in the other thread is indeed mine and has factory OC on the shaders at 1750. It is still slower than the 9800, but not by too much (about 29min vs. about 26min).


____________
141941*2^4299438-1 is prime!

For PPSieve, (the the PPSE sieve app):

• RAM speed doesn't matter much.
  
• RAM type doesn't matter much because RAM speed doesn't matter much.
  
• Compute capability minor revisions don't matter. (CC 2.0 gets a speed boost, but that's just because Fermi handles multiplications better.)
  

I should point out that there's a guy who's been working on primality proving with GPUs! He's gotten it to work with Mersenne numbers, but not like LLR with our numbers - yet. If he manages to make an LLR for GPUs, it will require double precision: that's compute capability 1.3 or higher.

Finally, how long can you wait? I've been working on an OpenCL version of PPSieve for ATI. It's going well. :) An ATI card might suit your heat and power requirements better. But I don't know how fast it would work yet. I don't have an ATI card to test on, just like I don't have an nVIDIA card to test on, largely because I've been putting off the decision until I see if ATI can work better.

There's a good chance that any hypothetical GPU LLR app wouldn't work on ATI.
____________