[gmx-users] Does GPU support ATI card?

[Justin Lemkul]

On 11/29/12 6:40 PM, Szilárd Páll wrote:
> Hi Justin,
>
> First of all, thanks for the feedback. I find in unfortunate that this
> information got public in the form of a "by the way". More official wording
> should follow. I wish it happened differently, but unfortunately things did
> not come together as we thought they would. Dragging along code that for
> the outsider seems supported, but in reality it is not properly maintained
> is not beneficial to anyone. Therefore, we decided to demote mdrun-openmm
> from "core" to "contrib" feature status. However, as we don't have a
> maintainer for the code yet, the official state is "not fully supported"
> contrib feature -- unless we find someone to help with at least minimal
> support for 4.6.
>
> I don't like to leave behind my own code, but the situation requires us to
> focus our efforts on other things. The plan is to include native GPU
> accelerated GB kernels in the next GROMACS release. While there is a good
> chance that the code itself will become available in a development branch
> rather soon, you will not be able to use this code in release version
> before the next release, 5.0.
>

If there's intent to create native GPU features within Gromacs, I'm totally fine 
with it.  Quality over time is certainly preferred to somewhat functional stuff now.

Further comments below...

> On Thu, Nov 29, 2012 at 7:54 PM, Justin Lemkul <jalemkul at vt.edu> wrote:
>
>>
>>
>> On 11/29/12 12:51 PM, Albert wrote:
>>
>>> On 11/29/2012 06:44 PM, Szilárd Páll wrote:
>>>
>>>> Hi Albert,
>>>>
>>>> That claim is false. The current mdrun-opnemm version only supports the
>>>> CUDA OpenMM plugin and even that is not fully supported as we don't have
>>>> enough resources to maintain the code.
>>>>
>>>> We will make sure to get the false claim of ATI GPU support removed from
>>>> the OpenMM website.
>>>>
>>>> Cheers,
>>>> --
>>>> Szilárd
>>>>
>>>> PS: we are looking for volunteers to maintain the Gromacs OpenMM
>>>> integration and in the lack of such people we will have to resort to
>>>> dropping OpenMM support post-4.6.
>>>>
>>>
>>> Hello Szilard:
>>>
>>>     thanks for kind comments. If there are not so much people use openMM,
>>> probably you can consider give it up. Currently I am using NVIDIA GTX590
>>> GPU
>>> acceleration with Gromacs4.6 without openMM support, and it works fine (I
>>> did
>>> some tests comparing with CPU running results for my system). If I were
>>> you, I
>>> will save my time and energy focus on current version and road.
>>>
>>
>
>
>>   This would be very unfortunate, I think, but I understand the need to
>> focus on the most feasible development, using the time people have to
>> commit to it, and perhaps priorities are elsewhere.  We invested heavily in
>> a few GPU workstations in recent years with the intent of doing a lot of
>> GPU-enabled implicit solvent simulations.  Without the GPU, the implicit
>> solvent code is severely limited without fixes to the parallelization on
>> CPU.
>>
>
> To be honest I am not aware of the limitations you are referring to, nor
> the current difference between CPU and GPU implicit solvent performance.
>
>
>> It seems that the same issues (i.e. one cannot use more than 2 processors
>> for a run) still persist thus far in the development of 4.6, with the focus
>> of using GPU's in a more
>
>
> I'm especially not aware of such a limitation and it sounds more like a bug
> than a feature.
>

Right, I filed a bug report a long time ago:

http://redmine.gromacs.org/issues/777

Berk applied a band-aid such that one can only use 2 processors for running on 
CPU.  In my mind, this remains a bug, but at least it is possible to run.  The 
failures are somehow related to constraints being split over multiple 
processors, and mdrun quickly exits with a fatal error if one tries to run using 
more than 2 processors.

I can certainly file a new Redmine feature request to make sure we consider a 
better resolution for the future.

>
>> generalized way, but without OpenMM, implicit solvent simulations (in my
>> hands at least) are not worth it.  Of course, running explicit-solvent
>> simulations on GPU is quite nice, but still, the point stands that the
>> implicit code is rendered largely obsolete without GPU.
>
>
> As you rely heavily on implicit solvent code in GROMACS, we could use some
> help from you. First and foremost, could you please report in form of a
> redmine issue and/or a mail to the dev-list the current issues you are
> having/you know of with both CPU and GPU implicit solvent code? We can
> continue the discussion on on those topic then.
>

I can certainly do that.  I suspect most of the problems (aside from the one 
described above) are related to OpenMM.  A good number of my simulations simply 
crash for no discernible reason.  I reported some a while ago and got no 
feedback, so I just basically abandoned that approach and have been waiting for 
the official 4.6 release to run explicit solvent stuff on GPU.  In any case, I'm 
happy to keep following up on such things.

-Justin

-- 
========================================

Justin A. Lemkul, Ph.D.
Research Scientist
Department of Biochemistry
Virginia Tech
Blacksburg, VA
jalemkul[at]vt.edu | (540) 231-9080
http://www.bevanlab.biochem.vt.edu/Pages/Personal/justin

========================================