personal PC configuration for GPU computing

Dear all,

I want to buy a new PC in purpose of scientific computation (Molecular dynamics) using CUDA. I am completely new in this era and I appreciate simple explanations.

My budget is about 1600$ for a case with out monitor. I have selected I7-2600k for CPU. for graphic cards I have access (existence in my country) to Geforce GTX 580/570 and 560Ti and for motherboard I have theses options; Asus p8p67 / p8p67 Delux / p8z68 Vpro / Subertooth p67.

I need guidance to select between these cards and the appropriate motherboards.

please guide me if its possible to add another card in future to run CUDA in multiple GPUs, for example is it possible to have 2 GTX 570 on these motherboards?

Thank you very much.

I don’t know anything about these motherboards, but you should be aware that the Intel processor you mention uses Socket 1155, which can only support a single GPU at full PCI Express 2.0 x16 bandwidth. If you put a second GPU into one of these motherboards, you will only have half the bandwidth available to either card. That might not be a problem for your application, but you should possibly check on that.

It depends on what program you use.

I can say that for Abalone speed depends on the frequency of memory rather than the number of processors. So 560 will be faster than 580

It’d go with the Z68 Chipset, since it is the most modern. The Asus board seems decent:
http://www.guru3d.com/article/asus-p8z68-v-pro-review/24

GTX580 is the fastest card single GPU card, memory-bandwidth and computation-wise.
@qNq: Sure the 560 is faster than the 580? Seems unlikely to me…

I use a 560 with 4400 MHz memory. If there is a 580 with the same frequency, it will be better. But it seems to me that no such card.

But! It concerns only the program that I mentioned. In other programs things may be different.

The GTX 560 has a smaller memory bus than the 580, so I don’t see how your overclocked 560 is a net win in performance. (It might be a win in price, however.) Annoyingly, NVIDIA has made two different “GTX 560 Ti” cards, so your GTX 560 card either has a 320 bit memory bus or a 256 bit memory bus. However, the GTX 580 has a 384 bit memory bus, so even running at a 10% slower clock, the overall bandwidth should still be bigger.

Is your code memory latency bound, and does changing the memory clock improve that somehow?

Yes. This is the main problem.