[MITgcm-support] MITgcm - CUDA architecture

[Samar Khatiwala]

Hi Stefano,

I've looked into this briefly with some colleagues from the Computer  
Science dept. here (hence my question some months
ago on this list about running MITgcm in single precision - no one  
responded!).

Essentially, what we found is that the cost of going back and forth  
between the fortran and GPU-native arrays/data-types
was nontrivial, and perhaps much too great to make this worthwhile.  
Admittedly, so far we haven't spend much time optimizing
performance or delving too deeply into the algorithm. If you do, it  
would be great if you could share your experience.

That said, I have an other application that makes extensive use of  
BLAS/LAPACK, and there the performance increase
was quite impressive.

Samar

On Jun 23, 2009, at 7:49 AM, Stefano Querin wrote:

> Hi everybody,
>
> I was wondering about the chance of porting the MITgcm model on  
> NVIDIA GPUs based platforms using the CUDA architecture.
> Those platforms seem to provide great performances.
> Does anybody have some experience about this topic, or does it  
> interest somebody?
> Here at my Institute, some colleagues would like to test these  
> platforms, and the MITgcm code could be a good case study.
> By the way, there is the will to do so (we don't have such a kind of  
> parallel system yet): I'm just looking for some further information,  
> if available.
>
> Thanks!
>
> Stefano
>
> _______________________________________________
> MITgcm-support mailing list
> MITgcm-support at mitgcm.org
> http://mitgcm.org/mailman/listinfo/mitgcm-support