Thank you very much for the reply David. This makes sense that cuda 9 is
mostly necessary to use new cards.
Keeping the same cuda (8.0) and gcc (4.8.4) on the DGX-1, I see a 5%
speedup for standard simulations with amber18 over amber16. My system is
not typical: 4 GPUs/run, charmm force field, 12 A cutoff, 800,000 atoms.
Still, I may as well post that there is indeed already a small speedup from
fully patched/updates amber16 to the base amber18.
Thank you again,
Chris.
On Wed, May 9, 2018 at 8:18 AM, David A Case <david.case.rutgers.edu> wrote:
> On Tue, May 08, 2018, Chris Neale wrote:
> >
> > I've compiled amber18 with cuda8 and gcc 4.x and I'm testing that
> > throughput now in comparison to the same setup with amber16. As I work
> > through these tests, I wonder if anyone has any data to indicate if cuda9
> > offers any improvement. Also, do more recent versions of gcc or openmpi
> > offer any improvements in amber18 performance?
>
> I'm not the best person to reply here, but will get started, and hope
> that others can correct what I say.
>
> Cuda version 9 is required for Volta based GPUs. More information is
> available in the "Supported GPUs" section at
> http://ambermd.org/gpus/index.htm. We are in the process of updating
> that page for Amber18. Note that Fermi cards (Hardware version 2.0) are
> no longer supported in Amber18, so hang on to Amber16 if you plan to use
> that (rather old, by now) hardware.
>
> Basically, MD performance depends a *lot* on the hardware you have, but
> very little on which version of cuda, gcc or MPI you use. Updated
> benchmarks are coming, but people that are really interested in small
> percentage changes will need to carry out their own tests, probably on
> the exact systems they are using.
>
> >
> > I guess I'm just trying to figure out if there is at the moment any
> > performance-based motivation to move from amber16 to amber18 (we already
> > bought amber18, I'm just hesitant to actually switch untill there is a
> good
> > motivation). Presumably the next gen of GPUs will give extra advantage
> with
> > amber18 over amber16 in any event.
>
> We hope/expect to issue a significant performance update for Amber18
> cuda code this summer. But right now, the difference in performance is
> small between Amber16 and Amber18, (assuming you have a fully-updated
> version of Amber16.) [Note: I myself am waiting for the updated
> benchmark suite results to be posted, since Amber16 vs. 18 diffs are
> probably not uniform across the four generations of cards (Kepler,
> Maxwell, Pascal and Volta) that we support.] If you have both versions
> of the codes, you can try some quick tests on your favorite systems and
> see what you find.
>
> There *are* significant addtions in Amber18 to the features that are
> supported: see http://ambermd.org/AmberMD.php.
>
> ....dac
>
>
> _______________________________________________
> 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 May 09 2018 - 09:00:03 PDT
