topic MKL DFT versus FFTW3.1 in Intel® oneAPI Math Kernel Library

MKL DFT versus FFTW3.1

Henneking__Stefan — Tue, 21 Mar 2006 21:48:01 GMT

It seems to me that the new version of FFTW (released this january) is faster (e.g. ~20% for 512 points complex 1d double_precision) than the MKL8.0 routines, at least on Itanium systems.
I just did some rough performance analysis with a code using either FFTW3.1 or MKL8.0. Does anybody have detailed data? The Intel page deals with the outdated FFTW3.0.1 only.

cheers
Stefan

Re: MKL DFT versus FFTW3.1

ttppo — Wed, 29 Mar 2006 22:52:17 GMT

512-point fft? FFTW seems slower than MKL8.0:
http://www.fftw.org/speed/Pentium4-3.60GHz-icc/

Re: MKL DFT versus FFTW3.1

ttppo — Wed, 29 Mar 2006 23:02:06 GMT

My God! MKL8.0.1 is even faster than AMD Core Math Library (ACML3.0) on Opteron.
Opteron is so slow! So poor the float-point preformence!
http://www.fftw.org/speed/opteron-2.2GHz-64bit/