If you mainly want to crunch here then a GTX660Ti is the card to get. If you want good CUDA performance at other projects I suggest you ask there. The GTX660Ti would do more work here, cost about the same and would cost less to run.

The last 1/3rd of the shaders can only be used under some circumstances by some applications and even then not entirely.

____________
FAQ's

HOW TO:
- Opt out of Beta Tests
- Ask for Help

As SK said.. GTX580 may still be the fastest here, but GTX6xx are far more power efficient. And a GTX660Ti rocks on POEM (mine was top host for a short time :D ).

MrS
____________
Scanning for our furry friends since Jan 2002

Note that there are two versions of the GTX560Ti:
The GTX560 Ti and the GTX560Ti 448.
The non-448 version only uses 2/3rds of its 384 shaders.

Which card you have?

A GTX580 would get through around 90% more work than the 384 version, but only about 15% more than the (210W TDP) 448 version, which can use all its shaders (cuda cores). This is down to Compute Capability (CC); the 580 is CC2.0 while the 560Ti (384) is CC2.1. All GF600 cards are the same (CC3.0), but this is more equivalent to CC2.1 due to shader access structure.

Note that when the GPU is idle it's still using some power. So the actual power consumption of the GPU is 130W plus the idle wattage.

A GTX 660 Ti should be significantly faster (perhaps 40% or more) than the 384 version and should use slightly less power; 150W TDP vs 170W TDP for the 560Ti (though there is more to it than that).

The GTX580 would do about the same amount of work as a GTX660Ti, and costs less, but would require about 30% more power to run.

I wouldn't bother 'upgrading' a GTX560Ti448 for ~15% more performance here.

For GPUGrid there is not much between the GTX660Ti and a GTX580, but I would probably get the newer card as it uses less power.
____________
FAQ's

HOW TO:
- Opt out of Beta Tests
- Ask for Help

This "1/3 of shaders unused" depends on the GPU-Grid code. NVidia implemented these additional shaders at relatively little cost (for them), but in order to use them the code must fullfil special conditions (topic "superscalar execution" - I coul give you further links if you're interested). This works to some extend in games, but not at GPU-Grid. Not that this applies to most chips/cards from the 400, 500 and 600 series. Only the flag ship chips have been exceptions: GTX480, GTX470, GTX465, GTX590, GTX580, GTX570 and "GTX560 448 cores edition".

If you wouldn't be able to properly reuse or sell your current GTX560Ti I wouldn't bother upgrading.

Although.. I think I can give you a good comparison point, since I've been running GPU-Grid on my GTX660Ti for the last 2 days. I've got WUs similar to the one you comleted in 8811 s (11250 credits). My system is idling at ~60 W, if I remeber correctly. Power draw running just GPU-Grid is 205 or 220 W (saw both numbers - maybe due to different WUs?). One CPU core and the memory controller are active, which requires about 30 W (as software tells me for my i7). The PSU is about 91% efficient, so all in all the GPU should use approximately (205-60-30)*0.91 = 105 W, or at maximum (220-60-30)*0.91 = 118 W.

At 1.19 GHz (maximum turbo stock GPU clock) these tasks require about 5700 +/- 100 s, with 50 MHz more I'm at ~5550 s at similar power consumption. I can run POEM at 1.33 GHz, but don't know yet about GPU-Grid.

One could also run the GPU at less than "full throttle". At maximum turbo the GPU voltage is pushed up to 1.175 V, which is nice for performance but not good for efficiency. Stock operation is approximately 1.0 GHz at 1.00 V. On should get a few 10 more MHz out of this. In this case power consumption should drop by ~40% with relatively little performance loss (~15%). You can't do this with Fermi GPUs since these usually already run at <1.00 V to keep efficiency somewhat in check.

MrS
____________
Scanning for our furry friends since Jan 2002

Where do I learn about the differences between Compute 2.0, 2.1 and 3.0?