In 2021, 2016's Pascal class GPUs (GTX 10x0(TI) etc) are old hardware IMO.
Maintaining GPUs before them means coding around their lack of virtual
memory support. We deprecated SM 2X a while back and it's not even clear in
CUDA 12 and beyond will even compile for SM 3 and SM 5 anymore as they are
both deprecated architectures.
But it's not a lot of work to keep supporting Pascal because Ampere ~=
Pascal + broken warp model + tensor cores. It is to keep supporting Maxwell
and Kepler. The upside of GPUs is they evolve much faster than CPUs. But
the downside of GPUs is also that they evolve much faster than CPUs. That
said, it's a great Faustian deal IMO.
At some point, tensor cores will work their way into MD. They already can
be wielded for AI models, but I have plans to optimize the NL builder with
them, a situation where approximate math is peachy keen. And one thing I
noticed is that by 1M atoms or so, the NL is built every other iteration so
it's an obvious target for improvement. If they ever support something like
accelerated tensor products and hadamard products, I could use them under
the hood for a lot of other things, but as of now, they're just 32x32 *
32x32 GEMM accelerators.
On Mon, Apr 5, 2021 at 4:57 AM Brent Krueger <kruegerb.hope.edu> wrote:
> Well, we still have quite a few GTX 780s in use that run great and are
> valuable. However, I completely agree that I can just continue to use
> AMBER20 on those systems. Our system is fully containerized now so I can
> likely continue to run old versions for a long time.
>
> I'm usually the one asking us to maintain support for old hardware, but
> agree that the support matrix could get a lot narrower in the near future.
>
> Cheers,
> Brent
>
>
> On Sun, Apr 4, 2021 at 9:19 PM David A Case <david.case.rutgers.edu>
> wrote:
>
> > On Sun, Apr 04, 2021, Scott Le Grand wrote:
> >
> > >CUDA 7.5 is a 2015 toolkit and SM 5.x and below
> > >are deprecated now. SM 6 is a huge jump over SM 5 enabling true virtual
> > >memory and I suggest deprecating support for SM 5 across the board. SM 7
> > >and beyond alas mostly complicated warp programming and introduced
> tensor
> > >cores which currently seem useless for straight MD, but perfect for
> > running
> > >AI models inline with MD.
> > >
> > >CUDA 8 is a 2017 toolkit. That's way too soon to deprecate IMO ...
> >
> > We currently (appear to) support CUDA versions 7.5 through 11.2, and SM
> 3.5
> > through 8.0. (All CUDA versions are not compatible with all SM's, but
> > allowed combinations are set by NVIDIA, not by us.)
> >
> > The argument for SM3 is that there are still K80's out there, although I
> > think this must be less and less true. There certainly are GTX 980's
> still
> > in use, and supporting SM5 would make sense if it doesn't cost anything.
> > But I'm fine with dropping support for SM5 if it helps us with code
> > development or maintenance. (If a future user really needs to run on
> older
> > cards, they can stick with Amber20.)
> >
> > The CUDA experts should make the decisions here, not me.
> >
> > ...dac
> >
> >
> > _______________________________________________
> > AMBER-Developers mailing list
> > AMBER-Developers.ambermd.org
> > http://lists.ambermd.org/mailman/listinfo/amber-developers
> >
>
>
> --
> _______________________________________________
> Brent P. Krueger (he/him/his)......phone: 616 395 7629
> Professor......................................fax: 616 395 7118
> Hope College...............................Schaap Hall 2120
> Department of Chemistry
> Holland, MI 49423
> _______________________________________________
> AMBER-Developers mailing list
> AMBER-Developers.ambermd.org
> http://lists.ambermd.org/mailman/listinfo/amber-developers
>
_______________________________________________
AMBER-Developers mailing list
AMBER-Developers.ambermd.org
http://lists.ambermd.org/mailman/listinfo/amber-developers
Received on Mon Apr 05 2021 - 11:00:02 PDT
