[Pw_forum] Using NVIDIA CUDA for blas with pwscf

[Axel Kohlmeyer]

On Mon, 5 Mar 2007, Detlef Schmicker wrote:

DS> Hi,
DS> 
DS> I am looking for cheap computational power to use pwscf.
DS> 
DS> Did anybody tried to use the GPU of NVIDIA graphic cards?

DS> The NVIDIA CUDA development kit offers blas and fft routines. But it is
DS> quite difficult to get information on performance of these GPUs,
DS> especially if you need much memory. 

depending on the application the performance increase can be up
to 30 (in ideal cases). i know from somebody who wrote a multithreaded
grid based simple poisson code across multiple GPUs with almost
200x speedup.

however, for BLAS/FFT the gain is less.

DS> The benchmarks show 2.4 times faster computation than with fastest
DS> woodcrest (intel) processors.
DS> http://rapidmind.net/sc06_hp_rapidmind_cpugpu_summary.php
DS> 
DS> Any information would be nice,

it is no point for pw.x and related codes, since 
the hardware does not support double precision 
operation. DP however, is _essential_.

you should not give up hope. the next generation hardware
is supposed to have double precision hardware and the API
is already there.

i'm quite curious to get my hands on that stuff, as soon
as it is available, too. :-)

cheers,
   axel.

DS> 
DS> Detlef
DS> 
DS> _______________________________________________
DS> Pw_forum mailing list
DS> Pw_forum at pwscf.org
DS> http://www.democritos.it/mailman/listinfo/pw_forum
DS> 

-- 
=======================================================================
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.