[gmx-developers] single precision, double precision hybrid GROMACS

Anton Feenstra feenstra at few.vu.nl
Fri Apr 7 09:41:15 CEST 2006


Mark Abraham wrote:
> Accuracy at "single" machine-precision level is already being swamped by 
> the assumptions inherent in the construction of the force-field - 
> uniformity of atom-typing, point-charge representation - and the MD 
> protocol - bond length constraints, temperature & pressure regulation, 
> electrostatics treatment, leap-frog integrator. There's probably more 
> effects that haven't occurred to me in the last two minutes! To quote 
> one of my doctoral advisers, once you have an MM force field 
> representing your physics, chasing too much accuracy here is like 
> "trying to jump over your own head."

Additionally, you should know that several critical calculations in the 
single precision version, are in fact already performed in double 
precision because they have been shown to be critical to the overall 
accuracy.

Also, not least important, you will have to consider writing mixed 
precision *assembly* inner-loops. Since these use the SSE (or similar 
'single instruction - multiple data') features of the CPU, I wonder if 
it is at all possible to mix precisions there and maintain performance.
If you do not write the assembly loops, you will most defenitely loose 
performance (possibly even more than a factor of two!).


-- 
Groetjes,

Anton

* NOTE: New Affiliation, Phone & Fax numbers (below) *
  _____________ _______________________________________________________
|             |                                                       |
|  _   _  ___,| K. Anton Feenstra                                     |
| / \ / \'| | | IBIVU/Bioinformatics - Vrije Universiteit Amsterdam   |
|(   |   )| | | De Boelelaan 1083 - 1081A HV Amsterdam - Netherlands  |
| \_/ \_/ | | | Tel +31 20 59 87783 - Fax +31 20 59 87653 - Room P440 |
|             | Feenstra at few.vu.nl - www.few.vu.nl/~feenstra/         |
|             | "If You See Me Getting High, Knock Me Down" (RHCP)    |
|_____________|_______________________________________________________|



More information about the gromacs.org_gmx-developers mailing list