I'm considering a new PC build, and various video cards for BOINC projects. This would be the first I've gone beyond CPUs for BOINCing.

Which would have the most effect on BOINC GPU productivity:

• Core clock speed
  
• Shader clock speed
  
• Stream processor count
  
• Memory size (512 MB vs 1 GB)
  
• Memory interface (64-bit vs 128 bit)
  
• Memory type (DDR2, DDR3, GDDR5)
  

I'm not the best expert for this, but I'll tell you what I know. I've heard that the important clock speed for BOINC is Shader clock speed and the number of stream processors. Note, however, that ATI/AMD will have a ton of stream processors while nVidia will use fewer CUDA cores; these numbers are not equivalent, so only use them to compare within a brand.

The memory size, interface, and type has an effect, but it is minor at best. This is more important when you are using the GPU for games and need a big frame buffer (ie for high resolution, lots of AA, etc.)

You should also consider whether you want ATI/AMD or nVidia. Some projects prefer nVidia's CUDA, such as PrimeGrid, while others work best with OpenCL or CAL on the Radeon cards, like Collatz or MilkyWay.

Happy crunching!

I have a couple ati's running. 320 stream processors. Just bought a cuda, 192 cuda cores, and it's running less points than the ati's, which were half the cost. I'm running collatz and milkyway, which I just read don't work well with cuda. How accurate is that statement, that they don't work well with these two projects?

CUDA is definitely better suited to projects like PrimeGrid and SETI.

Even though MilkyWay and Collatz's apps are both fairly basic derivations, I've been told and have learned through experience that ATi cards are simply better at running them. It's almost a 75% time increase between an HD 5850 and my GTX 460. Collatz's app takes about an hour for me but the mini app runs in 7 minutes. Again, faster on ATi.

To OP:

Both major manufacturers are quite different. As people have said, both types of card have their strengths and weaknesses. Often, one will excel where the other does not; for instance, NViDIA's CUDA is a very open development platform, whereas ATi's CAL/Brook is more difficult to code for and troubleshoot; ATi cards, though, often have higher shader clocks and more stream processors, which can significantly speed up some applications.

Memory definitely matters - not so much to BOINC and its productivity, but to your computer while you use it. Faster GPU = faster computer and less strain on the CPU, which has to feed and monitor it. Many GPUs nowadays have higher memory bandwidths and usually more than 512MB of RAM, unless you're looking at a budget card. If that's the case, then don't even bother. No point in dropping 80-100 dollars on a lower-end GPU that frankly you may not consider to be "worth it" after a while of crunching.

---

Ran some GPU Grid and Seti with my cuda. Slightly more points. ATI is definitely more bang for your buck, but of course ATI only runs three projects. This cuda I have is a fermi but I think I'll have to look at an even higher end one for the next purchase. Thanks.