[gmx-developers] GPU Code?

[Erik Lindahl]

Hi Axel,

On Mar 30, 2007, at 7:28 PM, Axel Kohlmeyer wrote:

> On Fri, 30 Mar 2007, Erik Lindahl wrote:
>
> [...]
>
> EL> > still be a good starting point.
> EL>
> EL> The GPU code isn't available for the time being; we're keeping it
> EL> internal to a collaboration with Vijay's group until we've  
> managed to
> EL> make it slightly more generally usable so we at least can  
> publish the
> EL> algorithms/implementation before giving it away to the  
> world :-) The
> EL> problem isn't "implicit solvent", but even getting  
> neighborlists to
> EL> work, and that it partly relies on non-disclosure stuff from  
> ATI that
> EL> we aren't allowed to redistribute.
>
> how about efforts towards writing a CUDA kernel using nVidia hardware?
>
> which is now available without NDA.
> http://developer.nvidia.com/object/cuda.html
>
> i've seen reports that people have managed to get over 700GFlOp/s
> out of a single box with three Nvidia Geforce 88XX cards.
> they also have a BLAS/FFT interface etc.

We're working on that too and are in touch with Nvidia :-)

A potential problem with Cuda is that it's a pretty high level  
interface. Both BLAS and FFT are really "simple" operations compared  
to the random access to memory of MD, not to mention the  
neighborsearching. One reason for going ATI was their CTM (=close-to- 
the-metal) low-level driver interface, which more or less exposes the  
assembly language...

Cheers,

Erik