Hey Scott,
JAC NVE shows the degradation nicely*. It has something to do with
register allocation for kernels exploiting warp-based programming, but NV
has been cagey about exactly what it is. It is *supposedly* fixed for CUDA
7.0, but even so, there really is no benefit in using 5.5, 6.0, or 6.5
(beyond newer is better(tm)) anyway even if there were no such bug. CUDA
7.0 is the next optimization target assuming said bug is really fixed.
Scott
*Only verified on GK110s and GK104s, may not manifest on Fermi (or as much).
On Thu, Sep 4, 2014 at 9:29 AM, Scott Brozell <sbrozell.rci.rutgers.edu>
wrote:
> Hi,
>
> I'm glad i asked; i assume that the 5% perf regression applies to
> non Amber software, eg, other MD codes ?
> Can you provide more details on the perf regression, compiler bug,
> etc.? Are there published benchmarks demonstrating the issues?
>
> This information would be useful to those of us forming compute policies
> and for communicating both up and down our chains of command.
>
> Thanks for the info and updating http://ambermd.org/gpus/
> with this advice.
>
> scott
>
> On Tue, Sep 02, 2014 at 02:02:43PM -0700, Ross Walker wrote:
> > To add some info.
> >
> > CUDA 6.5 is supported as of last week for AMBER 14 - see update.6
> >
> > The 5% perf regression over cuda 5.0 still exists in 6.5 though so 5.0 is
> > still the recommended version. I believe a fix for the perf regression is
> > planned for cuda 7.0.
> >
> > On 9/2/14, 1:26 PM, "Scott Le Grand" <varelse2005.gmail.com> wrote:
> > >The first rule of PMEMD Club is CUDA 5.0
> > >
> > >The second rule of PMEMD Club is CUDA 5.0.
> > >
> > >Were you thinking of building with 5.5, 6.0, or 6.5? OK. sure. We'll
> > >throw in a 5-10% unfixable perf regression (compiler bug introduced in
> > >5.5)
> > >for no additional charge.
> > >
> > >
> > >On Tue, Sep 2, 2014 at 12:50 PM, Scott Brozell <
> sbrozell.rci.rutgers.edu>
> > >> Are there particular reasons to build Amber14 with a specific cuda
> > >>version
> > >> among 5.0, 5.5, and 6.0 ? (aside from newer is better)
> > >>
> > >> When is Amber support for cuda 6.5 expected ?
> > >> The target machine has
> ...
>
> _______________________________________________
> AMBER mailing list
> AMBER.ambermd.org
> http://lists.ambermd.org/mailman/listinfo/amber
>
_______________________________________________
AMBER mailing list
AMBER.ambermd.org
http://lists.ambermd.org/mailman/listinfo/amber
Received on Thu Sep 04 2014 - 14:30:02 PDT
