Did you try CUFFT library? It runs pretty well on graphic cards.

http://docs.nvidia.com/cuda/cufft/

Didn't try it. I'm not sure if we have any Nvidia chips to run it on.

It is worth to try. With double precision you will calculate from 20 to 40 times faster depending on the enviroment, with single precision you will get from 6 to 8 times more. I guess in one-two years the difference will be even more impressive.

Well, FFTW seems to scale reasonably with the number of CPUs in HPC Systems (http://www.fftw.org/parallel/xolas.html). However, I doubt it to be true on a multi-core CPU. To my knowledge FFTs are mostly memory transfer limited, and this will limit the scaling significantly.

For the GPU FFT mentioned above. Getting a significant speedup depends very much on the size or the number of FFTs (many small ones) to be done. Also the performance is very much limited by the PCIe band-width. Including transfer from host- to device-memory and back does kill most of the advantage. Any good-scaling GPGPU algorithm does more than one operation on the same data directly on the GPU, before transfering back to host memory.

Hi frank,

We consider 2D FFT. It suggests at least N FFTs of length N. No problem with the large number of FFTs even if a single 2D FFT is calculated. What do you mean under small size of FFT? N cannot be more than 2^12 for 2D FFT. This is a limit for CUFFT library. Do you consider 4096x4096 as a small size? Or do you mean that for such length CUFFT will be less efficient? To some amount yes, but CPU made FFT will be also much less efficient because of far memory used.

Results of a test on my computer, 1048576 FFTs of length 1024:

CPU - 239.64 ms

GPU, double - 6.76 ms

GPU, single - 1.72 ms

Your last remark gives impression that only a small part of codes can be written on CUDA and calculated by GPU. Surely, it is not true. Any part of a program can be written on CUDA and calculated by GPU. And data transfer problem exists no more.

Sorry, 65536 FFTs of length 256 were calculated in the test. But times are given correctly.

Results of another test, 4096 FFTs of length 4096:

CPU - 365.42 ms

GPU, double - 14.55 ms

GPU, single - 2.42 ms

Dear Alexandre,

do the GPU timings include the transfer to the GPU device?

I was under the impression (c.f. http://blogs.mathworks.com/loren/2012/02/06/using-gpus-in-matlab/ ) that the issue of transfer overhead is an actual one. I'd be interested to correct this impression if wrong. Further, a GPU to me is a Vector processor. From my experience such a processor can only achieve a good performance (and outpace normal CPUs) if vectorization of sufficient length is possible. This is something many codes struggle to do.

I do not use matlab, I compile my C++ codes by Visual Studio Express. No data transfer is needed except for input/output operations or excessive use of memory.

To transfer 4096x4096 array of complex double my computer takes 85.78 ms. This confirms the statement that simple FFT with two transfers will take near the same time as single FFT by CPU. But when FFT is calculated much more times, which is often the case, my tests give a realistic estimation.