The problem isn't AMBER-specific at all. I've already demonstrated it
affects cuFFT. It probably affects *everything else* but if a codecbase is
either nondeterministic or solely uses floating-point math, it's hard to
detect. And this means such a GPU is fine for rendering the graphics
within games because who's going to notice a pixel color that's off by 1
part in a million?
So I fully expect the usual suspects to dismiss this as irrelevant...
But I disagree, the IPS repro I found last week shows ~1 in 1000 divergence
when it hits and that's more than enough to blow up a simulation. And my
point is that if we hadn't first seen the 1 in 1 million issues, I wouldn't
have gone looking for stuff like this. So take that, nondeterministic
single-precision peeps!
Scott
On Wed, Jun 19, 2013 at 10:50 AM, Daniel Roe <daniel.r.roe.gmail.com> wrote:
> Hi,
>
> On Wed, Jun 19, 2013 at 11:17 AM, Marek Maly <marek.maly.ujep.cz> wrote:
> > It's just a pity that the issues are here perhaps only Amber specific
>
> I don't think this is an Amber-specific problem; rather, it is because
> the Amber GPU code is fully deterministic that allows the problem to
> be found (Scott correct me if I'm wrong).
>
> -Dan
>
> --
> -------------------------
> Daniel R. Roe, PhD
> Department of Medicinal Chemistry
> University of Utah
> 30 South 2000 East, Room 201
> Salt Lake City, UT 84112-5820
> http://home.chpc.utah.edu/~cheatham/
> (801) 587-9652
> (801) 585-9119 (Fax)
>
> _______________________________________________
> 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 Wed Jun 19 2013 - 12:00:03 PDT
