Compiling shouldn’t be a problem, since compiling doesn’t require access to a GPU and recent CUDA toolkit distributions contain dummy versions of the libraries that come with the drivers.
Now running is a different story. I suggest first to try with the existing windows binaries, since they already contain the GPU package, so you can see if the GPU is properly configured and can be used at all.
The main difference would be if you need to use the KOKKOS package which currently cannot be compiled with the cross-compiler for the windows binaries.
Mind you that recent LAMMPS (patch) versions can also be compiled natively on windows using the Visual Studio community edition (available at no cost to individuals for personal use and with open source), however I don’t know if anybody has tried to compile either KOKKOS or the GPU package that way with CUDA.
Read more here: Source link