Right....the SPEs act like the stream processors on GPUs (the user  
kernel must
be in charge of pushing/pulling data in/out and be aware of SIMD  
issues).

I didn't know there was a port for PPC...that's awesome.

I'm not a AMD(nor ATI) GPU fan....the SDK seemed less mature than  
nVidia's when
I looked at them; but that'll change with OpenCL, right ;) ?  This was  
around the time they first
released their FireStream line.

Currently I do most of my computations on nVidia 8800's (Ultra and GT  
for PCIe 2.0x16)
and Tesla S1070 (I use 2-4 GPUs per Node).

I've been ignoring the Cell for a few years now because of the lack of  
support/tools from
other vendors and price points for GFLOPS vs development time;  
however, when I
recently investigated them again, it's now a option I would like to  
explore.

I'm currently in the process of migrating our Linux nodes to  
OpenSolaris because
the interrupt handler and network stack in Linux sucks (it doesn't  
scale the way
I need it to...(Open)Solaris does this very nicely) and DTrace is a  
must have
for me ;) ... I don't know what I would do without it now.  Thus my  
question about
OpenSolaris and PPC/Cell.

My main goal for this little experiment is to get OpenSolaris on PS3  
and test out
the Cell SDK and algorithm development and use proven tools like  
DTrace to
help with bottlenecks/optimizations.  Then move the processing to either
a Cell daughter card or  Cell blade(s).

Thanks for the reply

Carlos

On Mar 12, 2009, at 6:39 PM, Rayson Ho wrote:

> Currently there are 2 ways to "run" cell:
>
> First way is to run cell as a CPU -- the OS runs on the PPEs and the
> tasks (similar to threads) runs on the SPEs. Note that there are lots
> of limitations on what the tasks can do, and also note that the tasks
> do not have the ability to directly address the main memory.
>
> There is a port of OpenSolaris for PPC, and as PPEs implement the PPC
> ISA, only minor work is needed to run OpenSolaris on Cell or PS3.
>
>
> The second way is to use Cell processors on an extension card, and
> this is similar to a graphics card running GPGPU programs.
>
> However, why use Cell when a graphics card can do what you want at a
> cheaper price and offer better performance?? The latest AMD (ie. ATi )
> RV770 card can get you 1.2 TFLOPs of single precision or 200 GFLOPs of
> double precision FP performance... the next generation RV790 card is
> rumored to have way more (680 GFLOPs) of double precision performance.
>
> http://tech.slashdot.org/article.pl?sid=09/03/09/1838216
>
> Rayson
>
>
>
> On Thu, Mar 12, 2009 at 3:21 PM, Carlos Cardenas
> <carlos.cardenas at swri.org> wrote:
>> Has anyone done anything with Cell Processing and (Open)Solaris?
>>
>> I don't see anyone trying to port OpenSolaris to the Cell (like  
>> running on a
>> PS3 or anything) but I was
>> interested if anyone was using one of the following cards with  
>> OpenSolaris:
>>
>> http://www.mc.com/microsites/cell/ProductDetails.aspx?id=9434
>> http://us.fixstars.com/products/gigaaccel/
>>
>> If nothing else, is there any interest (within the community or  
>> within Sun)
>> in getting support for the Cell in OpenSolaris?
>>
>> I would like to use Cell in some R&D but don't want to run Linux if  
>> it's not
>> necessary.
>>
>> Carlos
>> _______________________________________________
>> hpcdev-discuss mailing list
>> hpcdev-discuss at opensolaris.org
>> http://mail.opensolaris.org/mailman/listinfo/hpcdev-discuss
>>


Reply via email to