[gmx-users] Performance of GTX 980 and 970

[Szilárd Páll]

On Mon, Sep 29, 2014 at 7:32 PM, Carsten Kutzner <ckutzne at gwdg.de> wrote:
> Hi,
>
> On 29 Sep 2014, at 16:27, Szilárd Páll <pall.szilard at gmail.com> wrote:
>
>> On Mon, Sep 29, 2014 at 2:46 PM, Mirco Wahab
>> <mirco.wahab at chemie.tu-freiberg.de> wrote:
>>> On 29.09.2014 14:31, Pappu Kumar wrote:
>>>>
>>>> I am wondering if anyone tested the performance of new GTX 980 and 970
>>>> cards and compared to 780/780Ti/Titan using the input systems given here
>>>
>>>
>>> As long as I don't have a card, I can only guess.
>>>
>>> But: until the appearance of reasonable benchmarks, you
>>> can look at the F at H-results:
>>>
>>> single:
>>> http://images.anandtech.com/graphs/graph8526/67747.png
>>
>> Note that that's the GB single precision benchmark's plot which shows
>> surprising large performance gap between 780 Ti and 980 - perhaps
>> there's something in the OpenMM GB kernels that favors Maxwell.
>>
>> This is the explicit solvent performance plot:
>> http://images.anandtech.com/graphs/graph8526/67746.png
>> showing ~25% improvement over a 780 Ti (still pretty impressive). This
>> is close to what I'd have guessed based on extrapolating the 750 Ti
>> performance.
> Wouldn’t one expect the 980 to be slightly slower than the 780 Ti
> due to its smaller single precision processing power?

You are referring to the decreased (theoretical) instruction
throughput per multiprocessor of by 1.5x, right? Considering only this
aspect, you are right, one would expect lower performance.

However, the Maxwell architecture seems to be quite a bit more
efficient than Kepler, in particular I think it has become much better
at instruction scheduling (as advertised). That's not such a surprise
as there was quite some room for improvement, I think. The theoretical
IPC of 7 on Kepler was like a dream that never materialized, even
codes with the right mix SP instructions would struggle to get above
an IPC of 5-6.

I have not profiled code on Maxwell extensively (yet), but what I know
from having compared our Ewald non-bonded kernel's peak throughput on
the 750 Ti vs 780 Ti is that the clock to clock* per multiprocessor
throughput is only ~15% higher on the 780 Ti which is much less than
the decrease in instruction throughput. Additionally, the GTX 980 has
and extra multiprocessor (16 iso 15) and higher base frequency than
the 780 Ti to compensate.


*Considering base clocks is not entirely correct as the stable boost
clock delta might differ.

--
Szilárd

> Carsten
>
>>
>> Also note that this is OpenCL performance which is inherently worse
>> than CUDA and the OpeCL-CUDA difference may not be the same across all
>> architectures!
>>
>> Cheers,
>> --
>> Sz.
>>
>>> double:
>>> http://images.anandtech.com/graphs/graph8526/67748.png
>>>
>>>
>>> M.
>>>
>>>
>>>
>>> --
>>> Gromacs Users mailing list
>>>
>>> * Please search the archive at
>>> http://www.gromacs.org/Support/Mailing_Lists/GMX-Users_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-users or send a
>>> mail to gmx-users-request at gromacs.org.
>> --
>> Gromacs Users mailing list
>>
>> * Please search the archive at http://www.gromacs.org/Support/Mailing_Lists/GMX-Users_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-users or send a mail to gmx-users-request at gromacs.org.
>
>
> --
> Dr. Carsten Kutzner
> Max Planck Institute for Biophysical Chemistry
> Theoretical and Computational Biophysics
> Am Fassberg 11, 37077 Goettingen, Germany
> Tel. +49-551-2012313, Fax: +49-551-2012302
> http://www.mpibpc.mpg.de/grubmueller/kutzner
> http://www.mpibpc.mpg.de/grubmueller/sppexa
>
> --
> Gromacs Users mailing list
>
> * Please search the archive at http://www.gromacs.org/Support/Mailing_Lists/GMX-Users_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-users or send a mail to gmx-users-request at gromacs.org.