[gmx-users] fftw vs cmkl

[Mark Abraham]

Dallas B. Warren wrote:
> Thanks Mark.
> 
> My stats on this, GROMACS 4.0.7 compiled with gcc, run on two quad-core Xeon processors with 16GB of RAM, 100,000 atoms, PME, 200,000 steps
> 
> fftw 3.2.1 - 20,027 seconds
> 
> Intel cmkl - 19,792 seconds
> 
> So about 1.2% faster.

IIRC icc does a better job on mkl than gcc. Of course you need to make 
sure the PME job is being dominated by the FFT part to be sure you're 
measuring a relevant speed difference. The output at the end of the log 
file reveals this.

Mark