[gmx-developers] performance 4.6.3 vs 5.0rc1

[Szilárd Páll]

PS: Note that I am only referring to the GROMACS kernels, generalizing
in the above manner to all CUDA codes would be just silly.
--
Szilárd


On Fri, Jun 27, 2014 at 4:27 PM, Szilárd Páll <pall.szilard at gmail.com> wrote:
> The source is me - and I develop the CUDA kernels so I do test
> compilers and drivers regularly and quite extensively. Feel free to
> verify the claim though, for reference here are a few numbers you
> should be able to reproduce ;)
> http://goo.gl/y9Kz4P
>
> Note that the performance bug has been reported to NVIDIA and resolved in 6.5.
>
> Cheers,
> --
> Szilárd
>
> On Fri, Jun 27, 2014 at 4:14 PM, Vedran Miletić <rivanvx at gmail.com> wrote:
>> 2014-06-27 14:29 GMT+02:00 Szilárd Páll <pall.szilard at gmail.com>:
>>> 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
>>
>> Szilard,
>>
>> do you have any source for that claim, or is this your experience? I
>> can only find a thread at NVIDIA forums [1].
>>
>> Regards,
>> Vedran
>>
>> [1] https://devtalk.nvidia.com/default/topic/745471/cuda-6-0-performance-drop-comparing-to-5-5/
>> --
>> 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.