[gmx-users] Comparing single with double precision runs

[Jussi Lehtola]

On Thu, 09 Jun 2011 17:30:10 +0200
Kalavera at gmx.net wrote:
> Dear Gromacs-users,
> 
> I compiled a single and a double precision version of gromacs 4.5.4
> (MPI enabled). The differences between single and double precision
> runs have been discussed in older threads (2005/2008) in this mailing
> list before. At least, I did not understand completely, which parts
> or calculations in an simulation are affected by my single/double
> precision choice. Some threads identified the equilibration step to
> be the most susceptible for differences. Is this true ? 

You might need to use double precision in minimization if your initial
structure is pathological (very large forces).
 
> With each version (single and double precision), I calculated the
> same simulation (same seeds). Which observables/energies/tests should
> I observe or conduct to find and evaluate the differences ?

Since you're in any case modeling a chaotic system, single and double
precision calculations might lead to completely different trajectories.
What you would need to do is sample over many runs to get the ensemble
right. You probably should observe just the observables that you are
interested in.

However, if you are using thermostats (or barostats), using double
precision is mostly just a waste of resources, since the thermostat
will just waste the extra precision.
-- 
--------------------------------------------------------
Mr. Jussi Lehtola, M. Sc.         Doctoral Student
jussi.lehtola at helsinki.fi         Department of Physics
http://www.helsinki.fi/~jzlehtol  University of Helsinki
Office phone: +358 9 191 50 632   Finland
--------------------------------------------------------
Jussi Lehtola, FM                 Tohtorikoulutettava
jussi.lehtola at helsinki.fi         Fysiikan laitos
http://www.helsinki.fi/~jzlehtol  Helsingin Yliopisto
Työpuhelin: (0)9 191 50 632
--------------------------------------------------------