I know of the web page for lammps benchmarks but these appear a bit out of date.
With the understanding that systems differ, would some users provide a few of their own results in #atomsteps/sec for LJ and Reaxff bench models, or perhaps better simply point to a link that contains such information. I just want to get a general relative handle of performance
I’m running 2x RTX 3090 GPU’s on 32 cores, 1 OMP AMD achieving
190 mas/s for LJ using just cuda
75 mas/s for LJ using kokkos and cuda
.5 mas/s for reaxff using kokkos and cuda
thanks
pb
I’m willing to run others if requested.
Hi there is some data here: LAMMPS - a flexible simulation tool for particle-based materials modeling at the atomic, meso, and continuum scales - ScienceDirect. It is a little out of date since ReaxFF has been optimized. I can also run some fresh benchmarks next week if needed.
Thank you very much.
Regarding fresh data, please do not go out of your way. Going through the reference you sent will keep me occupied for a while…
What spurred the original request was the vast difference in running LJ via gpu or gpu+kokkos as compared to the reaxx model - almost a 400 fold difference - and wondering if that is what I should expect in production runs.
Paul
1 Like