can I config hybrid potential with CUDA or GPU package

LAMMPS LAMMPS Mailing List Mirror
1

hi all
i use hybrid potential (eam -tersof - lj)
can I config hybrid potential with CUDA or GPU package?
if yes please explain how to do?
i use geforce 670 gtx
thanks

2

hi all
i use hybrid potential (eam -tersof - lj)
can I config hybrid potential with CUDA or GPU package?

in principle, yes. there may be restrictions.
you have to study the description of the individual
packages which pair styles they support and
how it can be used in combination with the
hybrid pair style.

if yes please explain how to do?

please have a look at the documentation.
that is what it was written for.

axel.

3

I doubt it - but Mike and/or Christian can comment
on whether one or more of the hybrid sub-styles can
run on the GPU.

Note that only the USER-CUDA package supports Tersoff.

Steve

4

My understanding is that you can only do this with the GPU package. The GPU package currently does not support Tersoff. The best you can do is reduce the run time by the percentage time spent calculating eam and lj. This is done by running eam and lj on the GPU at the same time Tersoff is running on the CPU.

To do this, use "force" and not "force/neigh" on the package gpu line. In the pair_style hybrid line, put eam and lj before tersoff and run with -sf gpu or suffix gpu or add "/gpu" to eam and lj on the pair_style line.

- Mike

5

hi
thanks for reply
i study this page
http://lammps.sandia.gov/doc/package.html
about GPU, CUDA package
in this page explain for hybrid i must use gpu just in force mode and say “GPU neighbor list calculation currently cannot be used with a triclinic box. GPU neighbor list calculation currently cannot be used with hybrid pair styles.”
but no thing about CUDA and hybrid
and i want simulate graphene sheet that need use tersoff or airebo
GPU and CUDA not support airebo
and tersoff just supported by CUDA.
thanks
manzoor

6

thanks a lot
i edit input file and use other GPU 285 gtx and compile arch=sm_13 :

newton off

pair_style hybrid lj/cut 8.0625 eam tersoff
pair_coeff * * tersoff SiC.tersoff C C
pair_coeff 1 2 lj/cut 0.010 3.225
pair_coeff 2 2 eam Cu_u3.eam

and use command -sf gpu but i receive this error:

7

Sorry, I forgot about the newton requirement. Currently you will be unable to use tersoff with the GPU package. Support for this might be added soon because we need 3-body for another model, but for now I think that there aren’t any options for your model.

  • Mike
8

Hi

the USER-CUDA package can not be currently used in hybrid mode but I am working on it for a Sandia project. What is it you actually want to do in hybrid? Right now you can use the GPU package with neighborlist construction on the CPU I think.

Christian

-------- Original-Nachricht --------

9

Hi
thanks for reply

i want simulate graphene sheet that need tersoff or airebo other potentials: lj & eam
in GPU package we need newton off but for use teresoff newton must on so i cant use GPU package for this problem.

manzoor

10

Hi
thanks for reply
i want simulate graphene sheet that need tersoff or airebo other
potentials: lj & eam
in GPU package we need newton off but for use teresoff newton must on so i
cant use GPU package for this problem.

rather than dwelling on what you _cannot_ do,
i suggest you can busy with what you _can_ do.
...and there is a lot.

have you made some tests running on the CPU?
are you sure that you what you want to simulate
is at all possible? how much acceleration do you
expect or need to get from the GPUs?
you can easily set up a correspondingly smaller
system and study many of the technical and simulation
related details, so that when you *can* run on
the GPU you are ready to move forward with your
project. also, i would venture a guess, that you can
be quite productive without GPUs if you plan your
simulations well. people have done this for decades
and with much more restricted computational
resources, that what we have available now.

the benefit of GPUs is that they allow a higher
throughput, but they are no magical wand that
make simulation happen in an instant. in my
personal experience, proper design of simulations
and workflow management are usually more important
than access to lots of computational resources.

cheers,
    axel.