[gmx-users] Normal Mode Analysis in Double-Precision

[Ran Friedman]

Hi Inon,

Check the files after EM. How many significant digits do you have? I
guess you wouldn't get a convergence to 5E-4 with single precision. From
my experience it may be worth to run several cycles of EM, say CG and
then L-BFGS to get better convergence.

Ran.


Inon Sharony wrote:
> I now performed:
>
> make distclean
> ./configure --disable-float
> make
> make install
> make links
>
>
> All executed without problems. Now I no longer have segmentation
> faults (thanks David!), however the energy minimization still does not
> achieve 1E-05 kJ / mole nm convergence (only as good as ~5E-04 for
> pentane). When I looked in the *.tpr file I saw that the coordinates
> are still given to three significant figures, meaning that the last
> two decimal places (fourth and fifth to the right of the decimal
> point) are always zero. I understood from Ran that this means the data
> is still in single-precision. How could this be?
>
> Can anyone please help me get proper, double-precision from my
> simulations?
>
>
> Thanks, and have a glorious week.
>
> Inon.