CPU vs GPU work units

[toTOW]

There have been discussion around GROMACS (the current CPU core) being ported to GPUs (hybrid or full GPU version), and there are some situations where it can run on GPU.

I'm quoting a GROMACS developer :

Virtual sites and all-bond constraints.
These are not supported by "all-gpu" version, unfortunately. We will, most likely, add virtual sites to the next release, but the all-bond constraints are out of the question for now.

All-bond constraints can lead to nonphysical behavior, especially when the system is very condensed. but for solvated non-membrane protein it should be fine. The reason why these options are not (yet) supported are purely technical: we are not a big team and we have to be selective in what we can do. I have got a couple of external request for virtual sites, so these will be most likely added to the new version.

[bruce]

Each analysis method covers all the basics plus a few things that may or may not be needed for a particular research study. toTOW quotes from the GROMACS perspective. OpenMM has similar internal tools that are not in GROMACS. (in fact, the less-than-stellar stability of OpenMM is a direct result of some new tools needed for the COVID Moonshot analysis and the bugs are still being corrected.)

Why do you ask? Allowing researchers to choose their preferred method seems to be working very well.

[mwroggenbuck]

I was curious.

I just wanted to know if CPU and GPU obtained different information. I guess I was thinking that maybe results from CPU were given to GPU for starting purposes or visa-versa. Now it is clear they both give the same information, but use different algorithms. However, for the purposes of FAH, it would not even matter if they used the same algorithm. This allows use of both CPU and GPU resources.

Don't read anything more into this. For some reason, I just assumed that different types of information were obtained from the different resources. That turns out to have been a bad assumption and made this thread quite long.

[bruce]

If the WU uses only features that are part of both GROMACS and OpenMM, they get the same results within the allowable round-off errors.

FAH uses the FAHCore_a7 (gromacs) for smaller projects and FAHCore_22 (OpenMM) for larger projects in order to use the hardware that Donors/volunteers happen to provide. In rare cases, the same analysis will be ported to both platforms.