[gmx-users] Supported features in Gromacs on GPU

[Andrew DeYoung]

Hi Mark,

Thank you so much for your kind response!  If you have time, may I please
ask a follow-up question?

In your response below, you said that three of the features that I asked
about (leapfrog integration, Nose-Hoover thermostat, and Ewald sum with
dipole correction) are, in fact, now supported on the GPU.  Were you
referring to version 4.6.x or to 5.0?  

Thanks so much!

Andrew DeYoung
Carnegie Mellon University 

> Hi,
>
> It seems that Gromacs support for GPU began with OpenMM (which works with
> Gromacs 4.5).  But starting with Gromacs 4.6, GPU support was native.
>
> I am interested in which features are currently implemented for GPU.  (I
> know that, given the staggering amount of code and years and years of work
> that have gone into Gromacs, it is a huge task to implement a lot of
> features on GPU!)  The page below lists some features supported in the
> OpenMM GPU-enabled Gromacs (i.e., 4.5):
>
>
>
http://www.gromacs.org/Documentation/Installation_Instructions_4.5/GROMACS-O
penMM#Supported_features
>
> But this page refers specifically to the older implementation (i.e.,
> OpenMM).  My question is, have any of the features unsupported in the
> OpenMM
> been implemented in the native GPU implementation (i.e., 4.6)?  The
> following page describes the native GPU implementation in general, but I
> don't see specific references to features supported:
>
> http://www.gromacs.org/GPU_acceleration
>
> I am particularly interested in these features:
>
> (1) ability to compile in double precision
>

No - see archive for post from Szilard Pall on this topic, but in short the
combination of needing a high-end card for actual good performance and
significant developer work that wouldn't be used by the core team is a bit
of deal breaker.


> (2) leapfrog integration (i.e., integrator = md; in OpenMM, only velocity
> Verlet is available)
> (3) Nose-Hoover thermostat (in OpenMM, only Andersen is available)
> (4) Ewald sum with dipole correction (i.e., ewald-geometry = 3dc)
>

Yes - the needs of an OpenMM implementation made life awkward for most
algorithms, but those limitations are history because we are no longer
working with OpenMM.


> (5) free energy calculations
>

Berk has been working hard for the upcoming 5.0 release, which should fill
in most or all of the blanks here - patch is in Gerrit awaiting us
organizing some test cases. I don't actually do such calculations, but if
you want to see if your needs are or might be met, please try out the code
from https://gerrit.gromacs.org/#/c/2803/ right away!

Do you know if there have been any changes in support for any of (1)-(5)
> above between versions 4.5 and 4.6?
>

The only thing the 4.5 and 4.6 implementations have in common is running on
GPUs and being called GROMACS.

But, again, I know that the developers are working hard and that it must be
> an extreme amount of work to implement things on the GPU.  I'm not
> complaining!
>

Sure. But if you're in the position to see if a test calculation does
something that seems correct, this is exactly the feedback we need during
the lead up to a release!

Cheers,

Mark