Low end machine for CUDA particle systems

Hi, I'm a physics student who uses matlab regularly as a hobby, and I want to use CUDA for particle systems. Should I build a mid-range new machine (i5/i7?) or take an older hex core AMD 1090T mobo/chip from a friend for free and put the money I save into the video card, taking me from the $200 to the $300/400 range on the video card? I'm thinking the CPU wont be as important. The slow part is two nested for loops which compute forces from every particle to every other particle using some field equations, so time scales exponentially with the number of particles. also, as long as I specify single precision, can I still do this with a single precision card without any performance hit? double is unnecessary and out of my price range.

Thank you! tldr; i5/i7 + $200 video card, or 6 core AMD Q1090 + %400 video card for particle systems with CUDA in matlab.

Answer 1

[Jan]

If you are a student and want to learn, buy the cheaper machine and invest time in optimizing the computations.

While faster machines might gain the double speed, improvements of the algorithms can accelerate a program by the factor 1000. You can find a lot of research for particle simulations. e.g. for galaxy collision. Some approximation techniques are very useful, e.g. considering the exact position of the near particles only and a less frequently updated field for the sum of all particles more far away.

Answer 2

[Jason Ross]

If you are going to use Parallel Computing Toolbox, you need a card with a compute capability of 1.3 or better (this means double). nVidia has a list here:

http://developer.nvidia.com/cuda-gpus

A GTX 580 would meet this requirement and your price constraint. Just make sure the motherboard has the right slot (PCI Express x16 x2), you have enough power (the card can max out at 200W by itself) and the case can fit it (it's a full size, double width card).

Another axis to optimize on is memory, which is relatively inexpensive these days, but if you can keep your calculations out of swap, you may be able to do it for a relatively small amount of money.