My Community

Hi team,

We need to run some calculations with dispersion corrections for elements past plutonium. The current D3 implementation (v6.5.0) is only parameterised up to plutonium. DFT-D4 is parameterised past plutonium and with minor modifications to the source code can be used with VASP for these elements, however is not compatible with GPUs – these are big calculations and GPU support is essential. Grimme did release an updated D3 version with parameters past plutonium (see https://doi.org/10.1039/D4CP01514B, implemented in https://github.com/dftd3/simple-dftd3/), but this was never included in VASP.

Are you able to update your D3 implementation to include elements past plutonium?

DFT-D4 GPU support would be ideal, but I believe this would require changes to the DFT-D4 code to make it compatible with Nvidia compilers as the fortran module must be compiled using identical compilers to VASP.

Cheers,
Ben

Dear Ben,

Thank you for your question.

Grimme did release an updated D3 version with parameters past plutonium (see https://doi.org/10.1039/D4CP01514B, implemented in https://github.com/dftd3/simple-dftd3/), but this was never included in VASP.

Actually we weren't aware of this extension. Thanks for bringing it to our attention. We will look into it.

DFT-D4 GPU support would be ideal, but I believe this would require changes to the DFT-D4 code to make it compatible with Nvidia compilers as the fortran module must be compiled using identical compilers to VASP.

I took a quick look at DFT-D4 code and I don't see any GPU related code, i.e., no OpenACC directive and no CUDA. So to be able to run DFT-D4 on GPUs it has to be ported first.

The DFT-D3 in VASP has not been ported to GPUs, but the workload related to the D3 corrections is rather small, so running a VASP calculation with DFT-D3 on GPUs should work just fine because all other parts are ported.

Best wishes,
Alexey

Actually, if DFT-D4 is parameterized for elements past plutonium you can directly use DFT-D4 in VASP, see the instructions here. As I said, neither DFT-D4 nor DFT-D3 is ported to GPUs, but the DFT calculation can be ran on GPUs and the vdW corrections can be computed on CPU without major overhead.

Hi Alexey,

Cheers for your reply.

My issue isn't so much running DFT-D3/4 on GPUs - as you point out the computational cost of running this part of the calculation is pretty negligible. Rather we'd like to use VASP compiled for GPUs, which also supports DFT-D3/4 for elements past plutonium.

DFT-D4 doesn't run on GPUs, but it is incompatible with the Nvidia fortran compiler (see https://github.com/dftd4/dftd4/issues/140). Since DFT-D4 must be compiled with the same fortran compiler as that used to compile VASP (https://dftd4.readthedocs.io/en/latest/recipe/vasp.html) I think that means it's impossible to use VASP on GPUs, with support for DFT-D4. If I'm wrong about that please do let me know.

Best wishes,

Ben

Sorry, I didn't understand your question at first. Now, I see what the problem is.

DFT-D4 doesn't run on GPUs, but it is incompatible with the Nvidia fortran compiler (see https://github.com/dftd4/dftd4/issues/140). Since DFT-D4 must be compiled with the same fortran compiler as that used to compile VASP (https://dftd4.readthedocs.io/en/latest/recipe/vasp.html) I think that means it's impossible to use VASP on GPUs, with support for DFT-D4. If I'm wrong about that please do let me know.

You are absolutely right. If DFT-D4 cannot be compiled with the Nvidia compiler, running VASP on GPUs with DFT-D4 will not work.

We have looked at the DFT-D3 implementation in VASP and it seems that extending the parameterization to the elements past plutonium would requires some work. We are putting it on our TODO-list, but I can't promise you that it will be done in the next release.