[CPMD-list] clearspeed acceleration for CPMD

[Axel Kohlmeyer]

On Thu, 25 Oct 2007, Joseph Hargitai wrote:

joseph,
 
JH> Anyone tried the cards with CPMD? If the card can accelerate dgemm 

no, i haven't tried. a while ago i asked about the price 
and then stopped thinking about it. since CPMD parallelizes
quite well with a good low-latency network, it is 
easier and more flexible to buy just mode cpus for
that money or a better network (if you are on gigabit).

JH> by 4.5x, can it reasonably accelerate CPMD?

well, it still doesn't say how fast you get the
data on and off the co-processor card. the latter
is a big problem for example with GPUs.
about five years ago, i made a comparison between
a netlib BLAS and ATLAS on CPMD on an AMD k7 cpu.
so the DGEMM in ATLAS was over 8x faster which 
resulted in "only" a 2x speed up of CPMD. to get much
faster from there will be even more difficult, since
now the major part of the time is spent in the FFTs.

if you had a code that doesn't parallelize (well),
then using clearspeed could be advantageous, but in
the case of CPMD i have my doubts. i still wouldn't
mind having access to one of those and doing some
serious tests and get hard numbers instead of having 
to gues, though... ;-)

cheers,
   axel.

JH> 
JH> j
JH> 

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