Amber does not need a lot of double precision arithmetic--I'm still
clearing out some occasional fp64 conversions when they don't, in fact, add
anything to the precision of the final result. What we've done is look at
the foreseeable range that users might need in mass production, and
tailored all of the arithmetic to maintain safe error bounds while
maximizing the fp32 and int32 throughput. The fractional space from real
space conversion, for example, is something I gave a lot of thought to.
You can't, for example, have an atom with charge greater than +/-21
electrons, for example, but no topology is going to have that without
deliberate tampering. Users who do have those situations will have to get
along with the 10x slower DPFP code, or the CPU version, but these fringe
cases are probably not something you'd want to scale up anyway.
Dave
On Fri, Nov 15, 2019 at 10:33 PM Shenglong Wang <sw77.nyu.edu> wrote:
> Thanks, Dave. With more readings, I found 0.509 TFLOPS for double
> precision with one card.
>
>
> https://www.microway.com/knowledge-center-articles/in-depth-comparison-of-nvidia-quadro-turing-gpu-accelerators/
> <
> https://www.microway.com/knowledge-center-articles/in-depth-comparison-of-nvidia-quadro-turing-gpu-accelerators/
> >
>
> Is it correct that Amber still needs very little double precision workload
> though most of the calculations are via single precision?
>
> Best,
> Shenglong
>
> > On Nov 15, 2019, at 10:16 PM, David Cerutti <dscerutti.gmail.com> wrote:
> >
> > Absolutely it works with these cards. I haven't been able to produce
> > benchmarks I am satisfied with because my RTX-6000 is not well powered,
> but
> > the calculations run fine. There is double precision on that card but
> not
> > nearly as much as something like a Titan-V. The difference between 6000
> > and 8000 is only memory, and the 6000 itself has gobs of it (24GB).
> > Doubling that to 48GB in the 8000 isn't necessary for any simulations
> that
> > Amber can run--other parts of the code will break before you make a
> system
> > large enough to hit that GPU memory limit.
> >
> > Dave
> >
> >
> > On Fri, Nov 15, 2019 at 11:36 AM Shenglong Wang <sw77.nyu.edu> wrote:
> >
> >> Does amber work well with NVIDIA RTX 6000 and RTX 8000? It seems they
> are
> >> single precision only without any double precision support.
> >>
> >>
> >>
> https://urldefense.proofpoint.com/v2/url?u=https-3A__www.nvidia.com_content_dam_en-2Dzz_Solutions_design-2Dvisualization_quadro-2Dproduct-2Dliterature_quadro-2Drtx-2D6000-2Dus-2Dnvidia-2D704093-2Dr4-2Dweb.pdf&d=DwICAg&c=slrrB7dE8n7gBJbeO0g-IQ&r=gI7YPDDMIksUN29Y65_UCg&m=bD7pGAUwuUqJAHWfWwWjyofSkdO-gDggwO6IFtCvAqs&s=pbQED7QU8vymG9Niyz-rap1w3YkjfW2JdcIjZyVhZRs&e=
> >> <
> >>
> https://urldefense.proofpoint.com/v2/url?u=https-3A__www.nvidia.com_content_dam_en-2Dzz_Solutions_design-2Dvisualization_quadro-2Dproduct-2Dliterature_quadro-2Drtx-2D6000-2Dus-2Dnvidia-2D704093-2Dr4-2Dweb.pdf&d=DwICAg&c=slrrB7dE8n7gBJbeO0g-IQ&r=gI7YPDDMIksUN29Y65_UCg&m=bD7pGAUwuUqJAHWfWwWjyofSkdO-gDggwO6IFtCvAqs&s=pbQED7QU8vymG9Niyz-rap1w3YkjfW2JdcIjZyVhZRs&e=
> >>>
> >>
> >> Best,
> >> Shenglong
> >> _______________________________________________
> >> AMBER mailing list
> >> AMBER.ambermd.org
> >>
> https://urldefense.proofpoint.com/v2/url?u=http-3A__lists.ambermd.org_mailman_listinfo_amber&d=DwICAg&c=slrrB7dE8n7gBJbeO0g-IQ&r=gI7YPDDMIksUN29Y65_UCg&m=bD7pGAUwuUqJAHWfWwWjyofSkdO-gDggwO6IFtCvAqs&s=GBXuK8e17tnRnwoZaqRst3bNytP_Y-MDA3xhh9jxNy4&e=
> >>
> > _______________________________________________
> > AMBER mailing list
> > AMBER.ambermd.org
> >
> https://urldefense.proofpoint.com/v2/url?u=http-3A__lists.ambermd.org_mailman_listinfo_amber&d=DwICAg&c=slrrB7dE8n7gBJbeO0g-IQ&r=gI7YPDDMIksUN29Y65_UCg&m=bD7pGAUwuUqJAHWfWwWjyofSkdO-gDggwO6IFtCvAqs&s=GBXuK8e17tnRnwoZaqRst3bNytP_Y-MDA3xhh9jxNy4&e=
>
> _______________________________________________
> 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 Sat Nov 16 2019 - 09:30:03 PST
