[Insight-users] ITK on Cell Broadband Engine

[Oleksandr Dzyubak]

Hi Daniel,

Just curious.
 From IBM guys I heard that PS3 (or blades with Cell BE) could be very 
productive because of:
1) Cell processors can be accessed (using IBM libs, of course) for 
parallel computing;
2) Blade memory circuits (eight in case of PS3) are distributed over the 
blade and
located in a very close vicinity of Cell processors (in contrast to the 
Intel CPU dime
where the memory is "grouped" and located "far away" from "CPU")
that shortens the data routing between Cell processors and distributed 
memory.

That's fine unless you are not overloading the onboard blade memories.
What if you are? Let say if I have to process 1 GB image.
In this case the code will request "external" (not on blade) memory
which actually can kill the productivity.

Do you have any experience in processing large (lets say over 100 MB) images
on PS3 (or Cell BE Blades)? Is a speedup still 10 (or even 60)?
What is the image size limit that will still keep a speedup high?

Regards,


Alex


Blezek, Daniel J., Ph.D. wrote:
> Hi Daniele,
>
>   We are still exploring how best to use the Cell.  In general, the PPU
> is a little underpowered, but will still run ITK very well, and the
> SPU's can be used for the "heavy lifting". 
>  
> I think you will have the best success modifying the MultiThreader to
> run on the SPU, this is the mode we currently work in.  However, the
> overhead of C++ burdens the limited local cache of each SPU, so we also
> write very tiny C routines for the processing.  In principle, this model
> fits very well with ITK, but requires re-writing inner loops.  You could
> try to run ITK directly on the SPU, but I think you will end up swaping
> overlays in and out.
>
> Best regards,
> -dan