Recommendations on NVIDIA cards for CUDA GPU computing

I use a Windows 7 PC with 8 i7 cores and 16GB RAM to run MATLAB and other simulation codes (i.e. FDTD software). I want to investigate how much of a performance boost we may get using GPU computing. I was wondering if any one had a recommendation for a configuration that might be good to start with. Nvidia recommends Tesla and Quadro GPUs for Matlab. However the price range between the entry level and the high performance parts is very large. We definitely don't want to build a high performance system yet because of the expense but also don't want to buy the cheapest since I have been told that it won't really increase performance over just using the CPU. So I would like to find the best bang for the buck. Can anyone make a recommendation?Thanks.

Answer 1

[Erik S.]

I agree with the previous answer, it depends on you application of course, but in general Nvidia Geforce cards offers alot of performance for their price.

[Mike]

Thanks. I started looking at the GeForce cards first but when I asked Nvidia they stated that Tesla and Quadro were recommended for MATLAB. I'm not quite sure how their product families are differentiated.

Answer 2

[Matt J]

You would need to consider how much RAM your computation consumes. You would want to have a graphics card with enough memory such that you don't have to be constantly shuttling data back and forth to the card.

Beyond that, there is no easy way to predict how well a particular card will do relative to the CPU because that depends very specifically on the structure of your computation, and how you intend to parallelize it. For the most part, though, high end cards like the Tesla are a lot more than what you need.

[Mike]

Thanks

[Kofi]

I have the same concerns. Trying to decide on the specifications for a new machine and wondering what benefits a GPU card will bring to my computations... mostly communications systems error rate Monte Carlo simulations that use few functions from the Communications Systems Toolbox and should parallelize well. I understand I'll have to get the Parallel Computing Toolbox to take advantage of the GPU.

Thanks

[Matt J]

@Kofi

Here is a list of operations that are GPU-parallelizeable in MATLAB,

http://www.mathworks.com/help/distcomp/run-built-in-functions-on-a-gpu.html