What is the GPU performance gain with Matlab?

Question

I'm thinking of asking my boss about a good GPU to do the calculations using the Matlab parallel processing dashboard. However, after some simple tests on a colleague’s computer, on the NVidia Quadro 600. I get a very poor performance increase (between 1x and 2x).

I am particularly interested in doing convolution on nd arrays.

Does anyone have similar experience doing calculations with help gpuArray, etc.?

What performance increase can be expected?

Which GPU model should I buy (less than $ 1000, hopefully less than $ 400)?

Answer 1

gpuArray MATLAB :

• .
• CUDA , ,
• (, Quadro/Tesla >= 2).
• MATLAB - R2011b , R2010b.

Walking Randomly: http://www.walkingrandomly.com/?p=3634. , - .. 5 10 , , .

( , , gpuArray arrayfun, 1500 , CPU: http://www.mathworks.com/products/demos/parallel-computing/paralleldemo_gpu_mandelbrot/paralleldemo_gpu_mandelbrot.html?product=DW)

Answer 2

Quadro 400 48 GeForce 480 480 GeForce 580 512 . , 10- , .

GPU MATLAB, , (CUDA) . Quadro FX5800 + GeForce 480, - GeForce 580. , GeForce. $$$ GF480 GF580.

, , CUDA, , -, 50 : http://www.mathworks.com/matlabcentral/fileexchange/20220 , MATLAB .

, ! . .

Answer 3

I am currently studying this topic. One thing I see is that low- and mid-range graphics cards are designed for gaming and are therefore optimized for FP32 performance. Their FP64 performance is far behind and may be what you see. The $ 1k GTX Titan seems to be the starting point for serious scientific computing.