How to use all GPU core

LAMMPS LAMMPS Beginners
1

Dear Lammps users and developers,

I just have compiled a GPU version of Lammps on a machine with Intel I9 (60 cores) + one graphic card RTX 5090. Reading some documentation, I understand that this 5090 have about 170 cuda cores. All my jobs started with mpirun

mpirun -np xx lmp -sf gpu -pk gpu 1 -in input >& output &

the maximum value of xx is 60, the number of CPU cores. I know there is a balance between CPU and GPU cores in the execution of the job, but I want to use all the GPU cores available. I read about split option:
split 1 → all GPU cores
When I use this option in the command line, the job never start complained about a parameter in the source file. How to do this? What I am missing?

Thanks

2

You always will. There is no way to control how GPU cores are used.
The way GPU acceleration works, all you can do is to make certain that you have enough work to do (e.g. enough atoms per GPU).

This controls the number of MPI tasks, with just one GPU, it makes no sense to go this far.
If you attach 60 MPI processes to a single GPU, each process gets 1/60th of a GPU to work with. That will cut into the usefulness. At the same time, you have 60 times the overhead of launching a GPU kernel.

Forget about it. This was something invented a very, very long time ago, when GPUs were far less powerful than today.