Re: GPU acceleration by using ATI video card

From: Aron Broom (broomsday_at_gmail.com)
Date: Fri Dec 14 2012 - 03:13:40 CST

Yes, I'd agree that OpenMM's speed advantage is most fully demonstrated
with implicit solvent (or even no solvent at all, provided there is also no
non-bonded cutoff). I didn't find it horrible for explicit solvent.

It's unfortunate that while GPU acceleration has actually been quite
successful in my mind, it was perhaps hyped up to an insane degree and so
was never going to be able to really meet expectations. It seems though,
that the 600 series (or at leas the 670 and 680) nVidia cards perform quite
well when most of the calculations are single precision, and so still offer
a good return on investment compared to a small cluster if you're certain
you're ok with running at that precision level (I think default AMBER for
instance is SPDP or SPFP or something that benefits quite well from those
cards).

So with any hope it will just be the GP-GPU specific market that dies off
and we'll all keep happily using consumer cards (albeit ones with larger
than average memory sizes).

~Aron

On Fri, Dec 14, 2012 at 4:04 AM, Axel Kohlmeyer <akohlmey_at_gmail.com> wrote:

> On Fri, Dec 14, 2012 at 9:49 AM, Aron Broom <broomsday_at_gmail.com> wrote:
> > I believe that at the moment this is true.
> >
> > For the time being, you can consider some alternatives. I believe OpenMM
> > will function with ATI cards, and hence, I suspect the GPU enhanced
>
> OpenMM only used to work well (and fast) for folding_at_home calculations,
> which are all implicit solvent jobs (well, at least that is what i am
> seeing on
> the machine that i am co-managing).
>
> > functions in GROMACS will also do the same. I thought also that LAMMPS
> > allows ATI cards?
>
> LAMMPS has two different GPU support modules (with different
> functionality).
> one of them can be retargeted to generate either CUDA or OpenCL code, which
> in turn makes it work on ATI/AMD. indeed, i have one node with 4x firepro
> cards
> and you can run LAMMPS on them. there is support for the usual "bio" force
> fields, but CHARMM support is (currently) lacking CMAP support.
>
> > OpenCL used to be fairly inferior to CUDA by all accounts (I'm speaking
> from
> > 2nd hand knowledge, so could easily be very wrong) but my impression is
> that
> > things are much more equivalent now, so perhaps both will be supported
> > eventually. From a GPU computing perspective this can only be a good
> thing,
> > as nVidia's pseudo-monopoly is not ideal.
>
> don't get your hopes up. considering the financial woes that AMD is in,
> they simply don't have the funds to support proper development of the
> drivers and utilities. they've been letting people go left and right.
>
> even nvidia's attempts to entice people to jump onto the GPU bandwagon
> these days have a ring of desperation to it.
>
> axel.
>
> >
> > ~Aron
> >
> >
> > On Fri, Dec 14, 2012 at 3:28 AM, dongsheng lei <laidongshengs_at_gmail.com>
> > wrote:
> >>
> >> Hello,
> >>
> >> I want to run GPU accelerated NAMD on my computer which has a ATI video
> >> card. However, according to the NAMD user guide and online information,
> GPU
> >> accelerated NAMD can only function with nvidia video cards, which are
> >> CUDA-capable. I want to know does it mean that we can not run GPU
> >> accelerated NAMD by using ATI video card in any case, currently?
> >>
> >> I heard that NVIDIA Contributes CUDA Compiler to Open Source Community
> and
> >> wish we can use ATI video cards to run NAMD with GPU acceleration in the
> >> future.
> >>
> >> Best wishes to you,
> >>
> >> Dongsheng Lei
> >
> >
> >
> >
> > --
> > Aron Broom M.Sc
> > PhD Student
> > Department of Chemistry
> > University of Waterloo
> >
>
>
>
> --
> Dr. Axel Kohlmeyer akohlmey_at_gmail.com http://goo.gl/1wk0
> International Centre for Theoretical Physics, Trieste. Italy.
>

-- 
Aron Broom M.Sc
PhD Student
Department of Chemistry
University of Waterloo

This archive was generated by hypermail 2.1.6 : Tue Dec 31 2013 - 23:22:51 CST