GPU support for solvent models?

Questions
#1

What, if any, solvent models allow GPU utilization? CMIRS apparently does not.

#2

The DFT part of the calculation can run on GPUs but not the solvation part.

#3

Thanks John.

I have tried a number of SOLVENT_METHOD choices. For the Iso-density SS(V)PE solvation model the SCF steps do not appear to use the GPUs. This is not the case for SOLVENT_METHOD = PCM THEORY = IEFPCM and CPCM) where the GPUs are active for the SCF Cycles but, as you stated, not for the actual solvation part of the calculation.

For the iso-density, the SCF is prefaced with:

Using 70 threads for integral computing
OpenMP Integral computing Module
Release: version 2.0, May 2017, Q-Chem Inc. Pleasanton

For the PCM the General SCF calculation program is used and GPUs are active.

For CPCM, is there any specific choice of parameters that will optimize speed without compromising accuracy?

Thanks.

#4

CMIRS uses iso-density SS(V)PE, which is separate code from PCM; this is explained in the manual: Q-Chem 5.1 User’s Manual : Chemical Solvent Models. For advice on which model to use, perhaps consider reading a review: https://doi.org/10.1002/wcms.1519

#5

Thanks. I just started looking into using implicit solvent so I was focused on the Q-Chem 6.0 manual.

Also, one of the first things I did was to lookup and read the review article you mentioned. I also viewed the tutorial you did for Q-Chem.

My choice of method is being driven by the article you mentioned and the utilization of the GPUs, since the system of interest is quite large (~1000 atoms and ~20000 basis functions).

That’s why I asked you about any performance optimizing methods that I could use.

#6

If the molecule is large then isodensity SS(V)PE, and therefore CMIRS, is likely to fail due to what Chipman calls the “star cavity” problem. For large PCM discretization grids, there are conjugate gradient solvers that get invoked automatically (though this is controllable, as discussed in the manual). However, with 20,000 basis functions, even a very large PCM grid may not be the rate-limiting step as compared to building and diagonalizing the Fock matrix.