Hamiltonian REMD, precision

GROMACS version: 2019, 2020
GROMACS modification: No


I am planning to run HREMD simulations using GPU acceleration. I am not absolutely sure if I can sacrifice accuracy for speed. Any comment is welcome.


Are you concerned about accuracy on the GPU? Why?

I have seen that the HREMD tutorial at the plumed site uses mdrun_d.

That rang the bell, since I have seen special methods (QM involved methods?), which require double precision. I just would like to know if HREMD needs double precision or not. I mentioned GPU just for completeness of the single/double issue.

The conventional wisdom is that only few MD processes really benefit from/need double precision. Free energy calculations and normal mode analysis are two examples. I don’t recall ever seeing a published REMD study that used double precision, because in effect it’s just plain MD with a small tweak to the energy function.