> I was wondering if there is a section in the amber manual that deals with
> the amber20 GPU input file set up specifically? Also, this error seems to
> be popping up intermittently - not easy to reproduce.
No there is not. For the fast GPU implementations, a mixed fixed precision
/ single precision model is the default and this means there is a limited
range in values for forces that do not blow the number format. If you run
double precision, it will be more tolerant or run on the CPUs where it is
double precision and more forgiving of numerical blow up (yet maybe not of
large force constants that can effect integration time step, etc). In
general, large force constants for restraints should be avoided since this
leads to a large range of numerical values (that blow up the precision
model). What we would question is why do you need a large force constant?
If you think of positional restraints with an Angstrom-kcal**2, even small
deviations lead to large numbers. Big deviations will make the integrator
fail, SHAKE fail, etc. I do not recall the original e-mail, but think of
very small force constants for restraints, otherwise dial back the
integration time step to much smaller.
That is my advise (ing).
> > >| WARNING :: The NMR distance restraint between atoms 9524 and 4574 could
> > >| be stretched to an unsafe degree. The run will proceed, but
> > >| overflow in the force accumulation is possible.
p.s. good that we have that warning - pre-emptive - reduce force
constants or time step or both.
--tec3
_______________________________________________
AMBER mailing list
AMBER.ambermd.org
http://lists.ambermd.org/mailman/listinfo/amber
Received on Wed Nov 02 2022 - 19:30:02 PDT
