[gmx-users] Benchs on gcc and pgi

[Erik Lindahl]

Hi,

On May 18, 2006, at 11:47 PM,
>
> Not wanting to press on this subject, that from previous  
> discussions in this list seems to be a somewhat "sensible" one: You  
> mean that the only and last thing that still need/should be  
> "converted" to make Gromacs run on GPUs (or somehow slacing the job  
> between the CPU and the GPU, as the tests from Folding at Home shows  
> to be the better option to the actual point) is "just" (not "just"  
> as a simples job, but as "the only and last one" job) the  
> neighborsearching? The "impossible" first, and "hard to get any  
> speed improvement" later situation has already grown to this point?  
> Those would be terrific news!

Well, Graphic card performance is increasing raplidly, but  
unfortunately it's a bit more complicated than that :-)

The main issue is that it is simply extremely complicated to work  
with neighborlists on GPUs, and huge amounts of data that has to be  
sent from the CPU to the GPU every 10 steps. It's not mainly a  
question of programming implementation, but rather that the  
performance drops completely when trying to do it.

Cheers,

Erik