Just a quick installation note. I was able to get vasp_gpu running for an older GPU card. The card has just 256 cores and 2GB of memory, but initial tests show some promise.
A changed to the GENCODE_ARCH parameter in the makefile.include was required:
GENCODE_ARCH := -gencode=arch=compute_30,code=\"sm_30,compute_30\" -gencode=arch=compute_35,code=\"sm_35,compute_35\" -gencode=arch=compute_20,code=\"sm_20,compute_20\"
Using CUDA Toolkit 8.0 and Intel Fortran compilers.
Calculations are gpu memory intensive. 2GB fills up quickly. I'm not able to use all 8 cores in my machine because each process spun off by MPI needs ~500 MB of memory on the gpu.