3 ms·

>I’m curious if anyone has a page with updated benchmarks for various sizes of FFTs on modern CPUs and maybe GPUs, testing whatever FFT libraries out there are

>I’m curious if anyone has a page with updated benchmarks for various sizes of FFTs on modern CPUs and maybe GPUs, testing whatever FFT libraries out there are still being actively maintained (including those which limit themselves to power-of-2 sizes) This is the most updated that I know of online (2014) https://github.com/JuliaLang/julia/pull/6193#issuecomment-61709788 https://github.com/JuliaLang/julia/pull/6193#issuecomment-61... If the array is large enough for a GPU, then yes a GPU version would rip it to shreds (not shown, but known). You can see the MKL advantage in there, but FFTW still performs pretty admirably. Everyone pretty much uses MKL or FFTW these days. I don't think there's as much development in this space because these are all pretty well-optimized now. Multi-process parallel for these is supposedly not difficult and can use a very similar strategy, but GPU really requires some different approaches and that's where the development is these days. Also, a lot of mathematicians have had a more recent focus on advancing the realm of spectral discretizations beyond FFTs, for example see Chebfun and ApproxFun, and things like FastTransforms.jl where the interesting this is non-uniformly spaced FFTs and similar transforms.

wyldfire 8y ago
Popular GPU FFT implementations are cuFFT [1] and clFFT [2]. IIRC cuFFT performs very well. [1] https://developer.nvidia.com/cufft https://developer.nvidia.com/cufft [2] https://clmathlibraries.github.io/clFFT/ https://clmathlibraries.github.io/clFFT/
jacobolus 8y ago
> for example see Chebfun and [Julia port] ApproxFun These are great technology which should be widely known and adopted. They are designed for working with functions defined on an interval rather than a periodic domain, and doing a bunch more with them. http://www.chebfun.org http://www.chebfun.org Ironically in the context of this discussion, these tools could get a whole heck of a lot faster with some dedicated effort from some low-level optimization experts.

wyldfire 8y ago

Popular GPU FFT implementations are cuFFT [1] and clFFT [2]. IIRC cuFFT performs very well. [1] https://developer.nvidia.com/cufft https://developer.nvidia.com/cufft [2] https://clmathlibraries.github.io/clFFT/ https://clmathlibraries.github.io/clFFT/

jacobolus 8y ago

> for example see Chebfun and [Julia port] ApproxFun These are great technology which should be widely known and adopted. They are designed for working with functions defined on an interval rather than a periodic domain, and doing a bunch more with them. http://www.chebfun.org http://www.chebfun.org Ironically in the context of this discussion, these tools could get a whole heck of a lot faster with some dedicated effort from some low-level optimization experts.