Help Choosing a GPU?

I am looking into purchasing a GPU for running a monte carlo simulation. My research involves simulating millions of photons moving through a detector system. A perfect problem for a GPU.

However, I have hit a wall with my lack of familiarity with GPU specifications. I have been looking at the Quadro and GeForce cards and have found quite varied numbers in CUDA processors, memory, compute capability and clock speeds. For the purposes of computing, which are the most important specs to consider? Also, do you have suggestions on processors that work better than others with Matlab?

We are looking to spend $200-400. Will the cards in this range give us a reasonable performance with regard to our simulations?

Thanks! Aaron

Answer 1

[James Lebak]

Obviously, it's going to depend on your problem, but the amount of memory on the card is a very important consideration. Estimate the size of your data set and make a guess at how many copies of the data might need to be in memory at once (just the input and the output? Or do you make intermediate copies as well)? If the cards you are considering can't hold this amount of data, then you may have to transfer data back and forth across the PCI bus, potentially erasing any performance gains you might otherwise have seen.

For the purposes of performance estimates, you might want to look at the GPUBench program on the file exchange. It includes performance results for MATLAB on different GPU hardware. You can find it here: http://www.mathworks.com/matlabcentral/fileexchange/34080

[Cedric]

Hi James,

Do you happen to know if there is an absolute best NVIDIA card for GPU computing, regardless of the price? I know that "best" has to be better defined, but sometimes there is a specific product (or products line) that has the preference of most people working in the field for a couple of months/years.. (?)

[James Lebak]

Well, the "best card for GPU computing" really does depends on what your application needs. Double-precision floating-point performance is really important for many applications. If you're looking for the maximum double-precision floating-point performance, then I believe NVIDIA's best offerings at this time are the GeForce GTX Titan in the consumer space and the Tesla K20 in the server space. NVIDIA provides some comments about how the GeForce and Tesla lines differ and some advice about how to use them for development and production in the following article: http://blogs.nvidia.com/2013/03/geforce-gtx-titan-cuda/

[Cedric]

Thank you James, this is exactly the type of insights that I was looking for! I am not on the market at the moment, but certainly within the next 6 months (I have currently GPUs with cap. 1.1 and 1.2 on my machines, which I am going to update), and your answer helps me starting categorizing existing hardware.