double precision CUFFT 2.3 is to slow

Accelerated Computing CUDA CUDA Programming and Performance
1

Hi @ all

I have compared the speed of the single precision fft with the speed of the double precision fft. (the parameter are nx 1024 and batch 3500 → the GPU calculate 3500 FFTs) Here are my results.

single precision: double precision

   3ms                        30ms

The double precision fft is 10 time slower than the single precision fft.

I used for the test a GTX 275.

Why is the double precision FFT so slow? (I used the cufft.dll version 2.3 for the test)

2

That sounds about right, there are physically 8 times fewer double precision units on the device.

3

I hope the new GT300 have more double precision units.