Hi,
I am trying to purchase a workstation with GPU card. There are many types of GPU cards with different properties. I am wondering if it is single precision or double precision simulations for LAMMPS, or we can
use some command to specify one of those simulations. Could someone give me some suggestions? Thanks very much.
Lili
Hi,
I am trying to purchase a workstation with GPU card. There are many types of GPU cards with different properties. I am wondering if it is single precision or double precision simulations for LAMMPS, or we can
use some command to specify one of those simulations. Could someone give me some suggestions? Thanks very much.
what is good and what is bad depends on very many factors.
several people have commented on that on this mailing list
in the past, so i would recommend that you consult the mailing
list archives to find those discussions and make your decision.
here are some simple general comments:
you do want to have a high-end GPU, but you don't
necessarily need a Tesla or Quadro card. the latter offer
mostly a higher quality hardware, but tend to be very expensive.
so much that, you may get the same price/performance
ratio with high-end CPUs as well.
if the fastest absolute performance in a single box is what
you want, no matter what the price, the Tesla K20 is the
hardware you want.
as for floating point, you can compile both GPU support
modules for single or double precision or a mix. all single
precision is the fastest, but errors may be significant for
larger and non-homogeneous systems. on most GPUs
a mixed precision approach is getting you close to the
level of accuracy of all double precision with the performance
close to all single precision. however, the ratio between
single and double precision capability changes between
different hardware generations and between consumer
grade and professional grade GPUs.
there are some LAMMPS GPU benchmarks on the web,
e.g. here: http://lammps.sandia.gov/bench.html#desktop
and here: http://users.nccs.gov/~wb8/gpu/bench.htm
axel.
Both GPU packages in LAMMPS (GPU, USER-CUDA) can
be compiled and run in either single, double, or mixed precision
on the GPU card. See the LAMMPS doc pages and Section_accelerate.html
for details.
Steve