Anon (no@email.com) on 7/30/10 wrote:
---------------------------

>I would disagree with you, for the reason that a single GPU can act in many ways
>like a stack of SSE+OpenMP+MPI, of course depending on your target performance,
>a single GPU may not be enough.. The current explosion of GP CPU cores is certainly
>causing changes in the balance here, but it was not long ago that anyone wanting
>more than a very few cores was using this kind of stack, and it is not 'easy'.
>
>

But the question is: is it easier or harder and GPGPU. The only in depth study I know of almost various HPC hardware types and software stacks is the work done by PRACE which showed a consistent trend of the SSE+MPI+OpenMP stack being consistently quicker to time to solution with significantly higher computational efficiencies.

A large repository of their work is at: http://www.prace-project.eu/documents/

Two presentations that highlight the results are:
http://www.prace-project.eu/documents/christadler.pdf
http://www.prace-project.eu/documents/02_wp8prototypes_hh.pdf

Over a lot of interesting data points and results.