[Pw_forum] Single precision FFTs on GPUs for QE ?
akohlmey at cmm.chem.upenn.edu
Tue May 15 18:04:01 CEST 2007
On Tue, 15 May 2007, Konstantin Kudin wrote:
i would suggest waiting until the end of summer as nvidia is
supposed to release cards that actually can handle double precision
(and will have accellerated FFT _and_ BLAS).
however, it should also be straightforward to test single precision
ffts with the existing codes, just change the FFTW code to call
the respective single precision subroutines. you have to convert
all data from double to single and back which will cost some
performance and i'm not so sure whether the resulting noise will
KK> Hi all,
KK> With all the hype surrounding the recent graphics cards and how fast
KK> they are supposed to be, the following question arises.
KK> If QE were to do everything in double precision, but the FFTs in
KK> single precision, will it hurt too much? FFTs kind of do few
KK> operations, so the error accumulation could be tolerable. Of course,
KK> another issue if the few digits of single precision are really enough
KK> With multicore cpus it is increasingly difficult to transfer the data
KK> between the boxes for FFTs, it sure would be easier (in principle) to
KK> send the data locally to the super fast (tm) GPU.
KK> Anyway, from the implementation perspective, FFTs are currectly
KK> reasonably well localized in the code, so in principle the hassle of
KK> using GPUs would be only local in terms of coding :-)
KK> ____________________________________________________________________________________Give spam the boot. Take control with tough spam protection in the all-new Yahoo! Mail Beta.
KK> Pw_forum mailing list
KK> Pw_forum at pwscf.org
Axel Kohlmeyer akohlmey at cmm.chem.upenn.edu http://www.cmm.upenn.edu
Center for Molecular Modeling -- University of Pennsylvania
Department of Chemistry, 231 S.34th Street, Philadelphia, PA 19104-6323
tel: 1-215-898-1582, fax: 1-215-573-6233, office-tel: 1-215-898-5425
If you make something idiot-proof, the universe creates a better idiot.
More information about the users