While this speed seems a little slow, the general performance seems to be
in the right ballpark. With a timestep of 0.5 fs (or 0.0005 ps), 0.25
ps/day corresponds to 500 MD steps per day, or one MD step every 172
seconds (about 3 minutes).
I haven't seen any quick-specific benchmarks, but terachem (another
GPU-accelerated QM code) has published some benchmarks for some older GPUs:
http://www.petachem.com/performance.html. The strategy for leveraging GPUs
in ab initio calculations is pretty straightforward, so the performances
are likely comparable. Of course these benchmarks were for pretty simple
systems (closed-shell organic compounds with small numbers of electrons per
atom and a rapidly-converging SCF). A B3LYP calculation can take
substantially longer for more complex systems (containing, say, transition
metals) or systems whose geometries or electronic structure is highly
strained, leading to slow SCF convergence (and therefore more iterations).
The end of the Amber mdout file should contain more detailed timing
information that could help you figure out where most of the time is being
spent.
HTH,
Jason
On Sun, Feb 20, 2022 at 8:12 PM xmgign.126.com <xmgign.126.com> wrote:
> Dear all,
> I'm using quick for QM/MM calculations and I had some problems.
> 1. My hardware is Tesla V100, but the speed of sander.quick.cuda is very
> slow. The number of atoms in the QM region is 46, and the overall number of
> atoms is 51956. The calculation speed can only run less than 0.25 ps/day.
> 2. The simulation broke down, but there is no error message. Is there any
> problem in my input file or the structure?
> Here is the input:
> qmmm md
> &cntrl
> imin=0,
> dt=0.0005,
> cut=8.0,
> ntb=1,
> ntt=0,
> ntpr=1, ntwx=50, ntwr=200, nstlim=2000,
> nscm=0, jfastw=4, ioutfm=1, ifqnt=1,
> /
> &qmmm
> qmmask='.1-27,3110-3114,6743-6753,29597-29599,',
> qmcharge=0,
> qm_theory='quick',
> qm_ewald=0,
> qmmm_int=1,
> /
> &quick
> method='B3LYP',
> basis='6-31G*',
> /
>
> Best regards,
> Qiao
>
>
> xmgign.126.com
> _______________________________________________
> AMBER mailing list
> AMBER.ambermd.org
> http://lists.ambermd.org/mailman/listinfo/amber
>
--
Jason M. Swails
_______________________________________________
AMBER mailing list
AMBER.ambermd.org
http://lists.ambermd.org/mailman/listinfo/amber
Received on Fri Feb 25 2022 - 07:00:02 PST
