Before running compile.sh make sure that you have:
- nvcc
- mpicc
To download and compile NVIDIA's official benchmark use compile.sh, providing the proper arguments.
./compile.sh -h
Usage: ./compile.sh -a <arch> -m <mpi> [-v <version>] [-cuda <version>]
-a Architecture: x86_64 | arm64-sbsa
-m MPI implementation: openmpi | mpich
-v Benchmark Version (default: 25.09.06)
-cuda CUDA Version 12 | 13 (default: 12)
Example:
./compile.sh -a x86_64 -m openmpi
For instance:
# Leonardo
# ml gcc/12.2.0 cuda/12.2 openmpi/4.1.6--gcc--12.2.0-cuda-12.2
ml gcc/12.2.0 nvhpc/24.5 hpcx-mpi/2.19
./compile.sh -a x86_64 -m openmpi