Parallelization efficiency of the GFN2-XTB in DFTB+ package

[DavidVrska]

Hello,

I'm working with the GFN2-XTB method embedded in the DFTB+ package (mainly because of periodic boundary conditions, and I'm also employing metadynamics, so I use the PLUMED package, which interfaces well with DFTB+).
 For the calculations (~200 atoms), I use GPU parallelization. I have found that the most efficient setup is ntasks-per-node=1 and --gpus=2; with more GPUs, the total computation time actually increases. The most effective CPU setup is slower comparing with GPU.

GFNn-XTB methods are supposed to be significantly faster than ab initio DFT, but in my case, I achieve only comparable speed (within the same order of magnitude).
Is it possible to speed up these calculations in some way?

Any suggestions or shared experiences would be highly appreciated. Thank you very much in advance!