Re: [AMBER-Developers] NVE vs NTP timings for latest gpu-tachyon code

From: Scott Le Grand <varelse2005.gmail.com>
Date: Tue, 6 Aug 2013 14:13:11 -0700

This why the Monte Carlo Barostat is crucial... I can't do much more for
traditional NTP - it needs more 64-bit oomph than NVE, twice as much in
fact, and that really hits consumeer GK104 and GK110 GPUs.








On Tue, Aug 6, 2013 at 1:34 PM, David A Case <case.biomaps.rutgers.edu>wrote:

> Taking the benchmarks from the Amber web site
>
> GTX780, cuda 5.5.22; driver is 325.15; code is current gpu-tachyon
>
> DHFR (NVE) = 106.9 ns/day; DHFR (NPT) = 78.6
> note that 106.9/78.6 = 1.36, which is a bigger penalty for NPT than we
> see on the Amber web site (ratio is generally about 1.23)
>
> Try the same thing on a C2050:
>
> DHFR (NVE) = 42.1 ns/day (vs web value of 43.7 for M2090);
> DHFR (NPT) = 36.1 (vs web value of 38.3 for M2090)
>
> ratio 42.1/36.1 = 1.16 (vs. web value of 43.7/38.3 = 1.14)
>
> So the "penalty" for running NPT vs NVE is bigger on the GTX780 than on the
> C2050 (identical code, drivers, cuda, etc.) Not sure if this means all
> that
> much, but the NBFIX test case I sent to Scott happened to be NVE, so maybe
> that is now super-optimized.
>
> So far, runs have been reproducible, but I only have one of these puppies,
> so am not too far along on the stress tests.
>
> ...dac
>
>
> _______________________________________________
> 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 Tue Aug 06 2013 - 14:30:03 PDT
Custom Search