Thank you for the helpful information. Looking into the mdout files, it
looks like the first step (minimization) does not run properly on the GPU
and the error propagates downstream - the energy at successive steps of the
minimization run remains the same. In contrast, on a CPU, the energy
decreases during the course of the minimization. So, I think this is one of
the main sources of the problem. Below is the input file for
minimization used:
minimization 01 - implicit solvent
&cntrl
imin = 1, maxcyc = 1000, ncyc = 100,
ntx = 1, ntc = 1, ntf = 1,
ntb = 0, ntp = 0, ntxo = 1, ioutfm = 1,
ntwx = 100, ntwe = 0, ntpr = 100,
igb = 1,
ntr=1,
restraintmask = ':1-X_LASTREC_RESIDNO_X & !.H= | :X_LIG_RESIDNO_X.CG',
restraint_wt=2.0,
&end
END
Any suggestions on what flags (if any) could be changed for single
precision runs on GPU? Or is the only way to tackle this problem, is to run
minimization on CPU first and then run the production on GPU?
Thank you for your help.
On Wed, Nov 9, 2022 at 11:24 PM Thomas Cheatham <tec3.utah.edu> wrote:
>
> > I tried to run the same simulation using double precision on the GPU and
> the output seems to be fine, whereas the
>
> Probably you mean CPU...
>
> > single precision seems to throw a bunch of errors (coordinate overflow
> etc.). Of course, the simulations with
> > double precision pmemd implementation runs a lot slower than what we
> would expect a GB simulation to run. Is there
> > a way/combination of parameters in the MD input file that could be run
> for the single precision implementation of
> > pmemd?
>
> No, and echoing what DAC said, often instabilities at the beginning of a
> simulation are due to a poor initial structure (which can be relaxed on
> CPU first). Another possibility that comes to mind is that it is possible
> to compile up different versions of the GPU code, for example the DPFP
> (double precision / fixed precision) which is slower but can handle the
> larger force variance better (likely, unless the initial structures are
> really poor).
>
> Regarding poor initial structures, we have seen peptides threaded through
> aromatic rings post initial relaxation and other distortions that lead to
> vdw overlap that is difficult to resolve. As per past posts, I often try
> initial minimization with electrostatics turned off to relax vdw (although
> this will not prevent a chain threaded through a ring).
>
> http://archive.ambermd.org/202101/0058.html
>
> --tec3
>
> p.s. delayed sending this since I wanted to send picture of peptide
> threaded through an aromatic from my lab, however we are not onsite that
> often
_______________________________________________
AMBER mailing list
AMBER.ambermd.org
http://lists.ambermd.org/mailman/listinfo/amber
Received on Fri Nov 11 2022 - 09:30:02 PST
