[gmx-users] Energy conservation issue

[Berk Hess]

>>
>>I love to hear this. Many people still consider achieving (nearly)
>>perfect energy conservation (for an isolated system) in single  precision
>>will be a big advantage. And I believe performing the integration  and/or
>>constraints in double precision will have negligible overhead, so I
>>would love to see this "hacking" can be made default inside Gromacs.
>
>The overhead would probably be insignificant on a standard x86 CPU,  but 
>Gromacs is also used on hardware where double precision has to be  
>simulated and is insanely expensive (embedded CPUs, graphics cards,  etc.).
>
>If at all possible I'd rather like to understand exactly _where_ we  loose 
>the accuracy and perform only that step on double, instead of  converting 
>entire arrays to double (Using bonds instead of  constraints will conserve 
>energy).

It' not that easy.
Making the constraint algorithm double precision is trivial, but has no 
effect.
I have tried making settle and the xprime coordinates in the update
double precision and that removes 1/3 of the drift.
I don't know where the other 2/3 come from.

Berk.