[gmx-developers] GPU Code?

[Erik Lindahl]

Hi Ansgar,

On Mar 30, 2007, at 5:42 PM, Ansgar Esztermann wrote:

> Hello,
>
> if I recall the issue correctly, one of the Folding at Home pages states
> that the gmx-based code that runs on GPUs has been contributed back to
> the gmx team. Is that correct?
>
> I have some time on my hands now and would like to look into writing
> GPU routines for gmx. I am not sure how much of the F at H code can be
> re-used (supposedly, it uses an implicit solvent model); but even if
> most or all of it would have to be re-written eventually, it would
> still be a good starting point.

The GPU code isn't available for the time being; we're keeping it  
internal to a collaboration with Vijay's group until we've managed to  
make it slightly more generally usable so we at least can publish the  
algorithms/implementation before giving it away to the world :-) The  
problem isn't "implicit solvent", but even getting neighborlists to  
work, and that it partly relies on non-disclosure stuff from ATI that  
we aren't allowed to redistribute.

It _will_ get there eventually, though; we have good hopes that the  
new generation of R600-powered cards will make neighborlists easier  
since they now finally support integer datatypes!

Cheers,

Erik