[lammps-users] Trying LAMMPS with GPU acceleration on NSF TeraGrid machines: NCSA Lincoln, TACC longhorn


is there anybody around on this list with access
to the NCSA lincoln machine or TACC longhorn, that
would like to try out the improved GPU acceleration
from the gpulammps project?

i can provide ready-to-use binaries for those machines
with code that will soon be merged into the mainline
lammps version and were you can help to identify
problems that still need to be addressed. we have already
run quite a few successful tests, so it should work well,
but having a larger pool of test inputs is always helpful
(and you get your work done faster).

compared to the existing code, the performance has been
much improved thanks to the hard work of mike brown (ORNL),
peng wang (nvidia) and others. exclusions and thus molecular
systems are now supported and a few additional pair styles.

current supported pair styles are:
gayberne/gpu, lj/cut/gpu, lj/cut/coul/cut/gpu, lj/cut/coul/long/gpu,
lj96/cut/gpu, cg/cmm/gpu, cg/cmm/coul/long/gpu.

the speedup per processor depends on the size of the problem
and the pair style, and ranges from about 10x to 20x for the
lj/cg family of pair styles and 40-60x for gay-berne models.

if you have any additional questions, please post them here.