Selim Sami

Right, thought so. LocalCoordinatesSite unfortunately won't work for what I have in mind. Could you perhaps point me at the right direction for implementing my own virtual site placement &...

That's very clear, thank you. I will give it a try and update here.

Hi @peastman, I've done the reference implementation and it seems to work well. Could you also point out for me what additions I would need to make for CUDA and/or...

Thanks a lot @peastman! That seems quite straightforward. I also noticed some virtual site code in [ComputeContext.cpp](https://github.com/openmm/openmm/blob/796ffaaa97decdf1593ed75afa5fbb21b6b32e9c/platforms/common/src/ComputeContext.cpp), is this something I should also change? A bit off-topic to this issue,...

@peastman, I have just tried with a conda installed openmm, and there are no changes in the performance. does openmm provide time accounting after a simulation? I did not manage...

Mostly because I wanted to compare it to my GROMACS implementation, I was eventually planning to switch to Langevin. But wow, I did not expect that much difference! I also...

Ah I see that explains it. The difference is between SCF and Lagrangian [is not supposed to be ](https://www.ncbi.nlm.nih.gov/pmc/articles/PMC4481176/)that big indeed. I find the [GROMACS](https://gitlab.com/gromacs/gromacs/-/blob/main/src/gromacs/mdrun/shellfc.cpp#L645) implementation rather simple and quite...

Oh that's great, thanks a lot for that. Looking forward to giving it a try. Another advantage of having a reasonably fast DrudeSCF is actually so that you can test...

That's interesting, but this might be connected to the DrudeSCF optimizer though. In GROMACS I am able to get < 0.1 kJ/mol/nm rms force in 12 SCF steps. And it...

Thanks I'll definitely give it a try. That was double precision. I just tried it in mixed precision and it can only go down to 0.04 kJ/mol/nm