[gmx-developers] Re: GROMACS 4.6 pre-release: new algorithms, parallelization schemes, & GPU acceleration

[Berk Hess]

On 03/05/2012 11:19 PM, Mirco Wahab wrote:
> Am 05.03.2012 22:42, schrieb Berk Hess:
>> All you need to add to make the gpu work, if you compiled with gpu
>> support is:
>> to your mdp file:
>> cutoff-scheme = verlet
>
> OK, lets see. I removed (from mdp) completely
>     vdw_type
>     coulombtype
>
I would use reaction-field, as that is probably closer to the Martini setup
and it will produce a smaller buffer and thus higher performance.
I think you would probably even need to use a shorter cut-off to match
the Martini interactions better. That I will discuss with the Martini 
developers.

Berk
> and inserted:
>     cutoff-scheme = verlet
>
> then, generated a new tpr,
>     ...
>     "Set rlist to 1.37 nm, buffer size 0.17 nm"
>     ...
> and
>
>     c:\> mdrun -nt 1 -nb gpu -v
>
> and *VOILA*, I can't believe this, it *does* work
> - Win7/x64
> - Cuda4.1
> - VS2010/SP1
> - fftw3f
> - gsl/gslcblas
> - libxml2
>
> A first test set of ~ 1 million particles (most MARTINI W,
> ~10,000 lipids) will run on a i7/920/(@3GHz) at about 11.5 GFlops
> with "*-nt 8*" (15.9 ns/day). On a stock GT-550 in the same
> machine (*-nt 1 -nb gpu*), it reaches  14.4 GFlops (23.5 ns/day).
>
> If I got some more time, I'll test the build against CUDA 4.0,
> according to Szilárds hint (possible performance drop in 4.1).
>
> I'm really impressed.
>
> Thanks & regards
>
> M.