• First, compile the code by typing make in src/. You will get the executables gpumd and nep in src/.

• Then, go to the directory of an example and type one of the following commands:

  • path/to/gpumd
  • path/to/nep

• By default, the nep executable will use all the visible GPUs in the system. This is also the case for the gpumd executable when using a NEP model. The visible GPU(s) can be set by the following command before running the code:

export CUDA_VISIBLE_DEVICES=[list of GPU IDs]
# examples:
export CUDA_VISIBLE_DEVICES=0 # only use GPU with ID 0
export CUDA_VISIBLE_DEVICES=1 # only use GPU with ID 1
export CUDA_VISIBLE_DEVICES=0,2 # use GPUs with ID 0 and ID 2

If you are using a job scheduling system such as slurm, you can set something as follows

#SBATCH --gres=gpu:v100:2 # using 2 V100 GPUs

We suggest use GPUs of the same type, otherwise a fast GPU will wait for a slower one. The parallel efficiency of the nep executable is high (about 90%) unless you have a very small training data set or batch size. The parallel efficiency of the gpumd executable depends on the number of atoms per GPU. Good parallel efficiency requires this number to be larger than about 1e5.

By default, the system is partitioned along the thickest direction, but one can overwrite this by specifying a partition direction in the following way:

potential YOUR_NEP_MODEL.txt   # use the default partition
potential YOUR_NEP_MODEL.txt x # force to partition along the x direction (the a direction for triclinic box)
potential YOUR_NEP_MODEL.txt y # force to partition along the y direction (the b direction for triclinic box)
potential YOUR_NEP_MODEL.txt z # force to partition along the z direction (the c direction for triclinic box)