Can not successfully compile VASP in GPU

[bhargabkakati]

No. I installed only one version of it (the latest version)

[bhargabkakati]

No. Unfortunately I don't have access to an older version. Do i need to install one? and how will that effect my already compiled codes such as QE

[bhargabkakati]

update :

Even though I am not able to execute test run using the vasp_ncl compiled using "makefile.include.nvhpc_omp_acc", I was able to successfully compile vasp with "makefile.include.nvhpc_ompi_mkl_omp_acc". But it is not very sped up. Even the CPU version compiled with "makefile.include.nvhpc_ompi_mkl_omp" runs 6-7 times faster (in 8 cores) than the GPU version. Is there any reason for that? or am I making some mistakes? Here I have attached the INCAR, POSCAR, KPOINTS, POTCAR and OUTCAR, and OSZICAR of both CPU and GPU runs.
Thank you.

[fabien_tran1]

Ok, it seems that you could compile the code on the GPU. There are several reasons that may explain a disappointing performance:

1) Your GPU model is RTX A6000, which is not among the recommended ones because it does not offer fast double-precision floating-point arithmetic (FP64) performance (see https://www.vasp.at/wiki/index.php/Open ... rt_of_VASP). Your model is described at https://www.techpowerup.com/gpu-specs/rtx-a6000.c3686 where "FP64 (double): 604.8 GFLOPS (1:64)" means that double-precision operations are 64 times slower than in single precision. A good GPU should have a "(1:2)" ratio (double slower than single by a factor 2). See also discussion at https://www.vasp.at/forum/viewtopic.php?t=19500#p26214

2) The benefit of running on a GPU depends on the calculation: size of the cell, method, used options, etc. You should make timing comparisons on various types of systems.

[bhargabkakati]

hello, thank you for the relevant information.

[fabien_tran1]

I have run your calculation (iron with spin-orbit coupling) on a machine with a GPU (NVIDIA A30) that has a fast double-precision floating-point arithmetic FP64. Similar as for you, the calculation on the GPU is much slower (by roughly 5 times) than on the CPU with 8 cores. So, this particular calculation is not a good case to show the power of VASP on GPU.

[bhargabkakati]

Thank you for sharing this information. Currently I am running different calculations on CPU and GPU to set up a benchmark, and so far, my CPU has been 2-3 times faster than my GPU.
CPU - Intel® Xeon(R) w9-3475X × 72 (hyperthreading on)