[gmx-developers] performance 4.6.3 vs 5.0rc1

[Szilárd Páll]

Thanks for checking! FYI CUDA 6.0 does have performance regressions,
but luckily on Kepler hardware the difference is quite small (<=5% in
the kernel and AFAIR only on CC 3.0), but on Fermi it can be up to
15%.

Note to self and to reviewers/documentation writers: I/we should note
this somewhere in the docs.

Cheers,
--
Szilárd


On Fri, Jun 27, 2014 at 8:39 AM, Mirco Wahab
<mirco.wahab at chemie.tu-freiberg.de> wrote:
> On 27.06.2014 01:50, Vedran Miletić wrote:
>>
>> 2014-06-27 0:51 GMT+02:00 Mirco Wahab <mirco.wahab at chemie.tu-freiberg.de>:
>>>
>>> gmx 4.6.3,  VS 2010, CUDA 5.5  =>  8.363 ns/day
>>>
>>> gmx 4.6.5,  VS 2012, CUDA 6.0  =>  6.590 ns/day
>>> gmx 5.0rc1, VS 2012, CUDA 6.0  =>  6.604 ns/day
>>>
>>> Maybe there are some significant "improvements"
>>> hidden in the CUDA 6 interface?
>>
>>
>> How about VS 2010 vs VS 2012 without GPU?
>
>
> If it's late at night and it's already impossible
> to keep both eyes open the same time, one should
> stop posting performance figures. Sorry for that.
> Here's the solution ...
>
> Large system, MARTINI-Vesicle (2.4x10^6 atoms):
> -----------------------------------------------
> Windows, gmx 4.6.3,  GTX-780,  VS 2010, CUDA 5.5  =>  8.363 ns/day
>
> Windows, gmx 4.6.3,  GTX-660Ti, VS 2010, CUDA 5.5  =>  6.610 ns/day
> Windows, gmx 4.6.5,  GTX-660Ti, VS 2012, CUDA 6.0  =>  6.590 ns/day
> Windows, gmx 5.0rc1, GTX-660Ti, VS 2012, CUDA 6.0  =>  6.604 ns/day
>
> Therefore, the 8.36 vs 6.61 figure represents the gain
> from a GTX-780 compared to a GTX-660Ti (which I missed
> when I studied the old logfiles). CUDA 6 does not lead
> to any performance degradation.
>
> Regards & sorry for the confusion
>
> M.
>
>
>
> --
> Gromacs Developers mailing list
>
> * Please search the archive at
> http://www.gromacs.org/Support/Mailing_Lists/GMX-developers_List before
> posting!
>
> * Can't post? Read http://www.gromacs.org/Support/Mailing_Lists
>
> * For (un)subscribe requests visit
> https://maillist.sys.kth.se/mailman/listinfo/gromacs.org_gmx-developers or
> send a mail to gmx-developers-request at gromacs.org.