These example scripts can be run with the USER-CUDA package, assuming you built LAMMPS with the package and the precision you want.

Note that these benchmark problems are identical to those in the examples/gpu directory which use the GPU package.

You can run any of the scripts as follows. You can also reset the x,y,z variables in the command line to change the size of the problem.

With the GPU package on 1 GPU:

lmp_machine -sf gpu < in.gpu.melt.2.5 mpirun -np 8 lmp_machine -sf gpu -v x 6 -v y 6 -v z 6 < in.gpu.phosphate

With the GPU package on 2 GPUs:

mpirun -np 4 lmp_machine -sf gpu -pk gpu 2 tpa 8 < in.gpu.melt.5.0 mpirun -np 12 lmp_machine -sf gpu -pk gpu 2 < in.gpu.rhodo

CPU-only:

lmp_machine < in.gpu.melt.2.5 mpirun -np 4 lmp_machine < in.gpu.melt.5.0 mpirun -np 8 lmp_machine -v x 1 -v y 1 -v z 2 < in.gpu.rhodo

Note that with the GPU package you can have more MPI tasks than the number of GPUs (both per node).

Also note that when running the in.gpu.melt.5.0 problem on the GPU, which has a long cutoff, the package gpu "tpa" setting should be > 1 (e.g. 8) for best performance.