For a research project on solubility I am doing free energy calculations on crystals and the electrostatic stage is a significant bottleneck as perturbed non-bonded forces cannot be calculated on the GPU.
What are the technical issues in the way of GPU calculations for perturbed non-bonded forces?
The free-energy kernel has many different branches and a lot of logic and is therefore more difficult to port to the GPU.
Also note that the PME workload nearly doubles, but that can be still be done on the GPU.
Which version of GROMACS are you using? The 2022 release introduced SIMD acceleration of the free-energy kernel, speeding it up by a facto 4 to 8. So now the non-bonded perturbed pair should no longer be limiting performance, unless you have a very weak CPU of are perturbing a significant part of the system.
I’m using 2023.1
For the crystal free energy, the entire crystal needs to be perturbed. In my case, it’s almost 20,000 atoms being perturbed. Have I missed an option somewhere? I just did a standard CUDA install.
Ah, that’s an unusual setup. Then all the cost is in perturbed interactions. Then the GPU will only compute the PME mesh part.