Support for GPU for acceleration?

[Sherwood]

Currently I saw the gmx_mmpbsa (linux version) is using sander to run gbsa/pbsa, which seems to the CPU version of Amber tools, so for guys who also have purchased Amber20, is it possible to use the cuda version of Amber such as pbsa.cuda or pmemd.cuda in gmx_mmpbsa to get faster results?

[gmx_MMPBSA]

Not at the moment, but we are planning its implementation (https://github.com/Valdes-Tresanco-MS/gmx_MMPBSA/issues/200). This does not require the Amber commercial license, since pbsa.cuda is free in AmberTools. pmemd does not support parsing (imin=5), so gmx_MMPBSA will not require it.
The implementation of pbsa.cuda may take some time due to the change it supposes. This implementation requires a hybrid computation scheme between MPI and CUDA, combining CPUs and GPUs. The Amber documentation says that they are working on this development, so it could speed up its implementation.
I imagine that if you have the Amber license, then you are doing the simulation with pmemd. In this case, we are also working on the Amber implementation for the calculations. I mean, the current version of gmx_MMPBSA, only supports Gromacs, however, we are working to implement official support for Amber and NAMD, since many users want to use the gmx_MMPBSA_ana analysis tool. 
We look forward to joining the Amber project as collaborators when we officially support it. Stay tuned for our updates, because our roadmap is very exciting.

Cheers!

Mario S.

[Sherwood]

Thank you Mario! any estimate when 2.2 release going to be available, we love to be a beta tester :)

Yes, we have a license of Amber and we do use pmemd.cuda to run simulations, so what are the additional features of gmx_MMPBSA vs the pbsa/pbsa.cuda came out of the AmberTools? the gmx_MMPBSA_ana is indeed great and sounds like it is not able to directly load Amber pbsa output (yet)?

[Mario Sergio Valdes]

Thanks, we'll keep that in mind. gmx_MMPBSA is growing considerably and help is welcome!

gmx_MMPBSA is based on MMPBSA.py as we mention in the documentation and the paper, however, we have implemented many functions and features that make it superior.
We have focused on providing as many methods as possible and on the implementation of all the features that Amber provides (which are many). For this, we have had to restructure a large part of the code, especially the integration of the input with the calculations and the output with our analysis tool. Currently, gmx_MMPBSA easily implements all possible options for any supported calculation type, without the need to generate/modify mdin files. Additionally, we have modified various usability aspects and definitions, as well as implemented new radii sets in ParmEd. All of this may not be very noticeable to the user, but it entails a considerable effort for its implementation and testing.
The main difference between pbsa.cuda and the current implementation is the calculation scheme. sander uses only CPU and computes both MM and PBSA components, while pbsa.cuda only computes PB and ELE components. This means that we have to implement a hybrid scheme to compute both at the same time (sander for MM and pbsa.cuda for PB), which would be completely new. The distribution of processes in CPU is easy because it is managed by MPI, however, it is not like that for GPUs, for which we would have to do several tests for its implementation. Also, pbsa.cuda does not support MD, unlike sander. The implementation of the GBNSR6 model planned for version 1.6.0 has a similar structure, so it will be the starting point.

No, gmx_MMPBSA_ana only supports the output of gmx_MMPBSA and as I mentioned, it has undergone considerable changes compared to MMPBSA.py. But we hope to officially support Amber soon.

Mario S.