[Beowulf] Rackable / SGI

Vincent Diepeveen diep at xs4all.nl
Thu Apr 2 19:23:16 PDT 2009


Yes a sad day,
especially considering they were supporting shared memory,
and i can just compile my chess engine, modify 2 settings and
it works right out of the box there at up to 2048 processors,
no problem.

MPI is a lot slower and would require a huge rewrite of my code.

Of course you can complain about 100 different things,
but in the end it is intel messing up with itanium.

Too late, too slow, too expensive, too little cores,
too outdated process technology, too complicated to port software to,
too intel c++ compiler dependant for performance,
too little attention given to integer performance,
too testset L3 size based approach,
too low clocked because of the outdated process technologies,
too weak engineering teams (as compared to the genius guys
who designed core2 and opteron).

There is still entire, once very expensive, itanium machines idling  
everywhere,
sometimes 1 scientist runs a RAM intensive program on it that hardly
eats the cpu's.

As i posted at RWT, initially a single partition altix3000 series of 64
processors (note bigger partitions allowed), had a cost of 1 million  
dollar.

Already half of that was just cpu costs, as the 1.5Ghz itanium2 had a  
cost
initially of 7500 dollar a piece.

That's if you could get it for that (who buys 1000 at a time?).

That my government ordered the 5+ times cheaper 1.3Ghz versions already
is very selfexplaining the fact that the cpu was simply too expensive  
for its
performance.

Of course if just 64 cpu's is already half a million, SGI had no  
choice but to
get a similar profit margin. On hardware that's not too massively you  
need
simply such a margin.

If processors would've been 50k dollar for 64 of them @ 1.5Ghz, then  
of course
it is possible to consider charging 50k profit on top of that, which  
would have
made the itanium2 really attractive platform.

Of course SGI never could deliver cheap, except for a few machines  
where intel
basically didn't earn on the cpu's, like that NASA machine. Typically  
those machines
got ordered.

Without that machine, SGI might have sold nearly nothing.

Vincent

On Apr 1, 2009, at 2:56 PM, Joe Landman wrote:

> Kilian CAVALOTTI wrote:
>> On Wednesday 01 April 2009 13:58:22 John Hearns wrote:
>>> http://www.sgi.com/company_info/newsroom/press_releases/2009/ 
>>> april/rackable
>>> .html
>>>
>>> Not an April Fools.
>> Whoa. That, and IBM eating Sun, the market is really shrinking...
>
> Read the note from SGI.  This was done as a bankruptcy filing this  
> morning, and not all liabilities are going over.  That is, it is an  
> asset purchase.  Creditors may still object, and force this to be a  
> chapter 7 filing (sell the thing bit by bit).  I had been wondering  
> if they had made their $5M debt service payment on friday, and  
> whether or not their creditors believed them to be in default.  The  
> speed of this and its construction, have a feel of a "rush" job ...  
> last minute to stave off a complete shuttering and sale.
>
> I sympathize with my friends and former co-workers at the  
> company ... this is a whopper of a layoff ... with likely no  
> severance (SGI was ~3x Rackable's size w.r.t. people, and I would  
> bet only a small fraction will follow the asset sale, if Rackable  
> makes them an offer).
>
> Sad day.
>
> As for the market shrinking, well ... this does happen.  We  
> (Scalable) are happy to help customers out (service/support on  
> existing systems, new clusters/storage), as is Don's company  
> (Penguin), as are other companies here.  Some of us are doing well  
> and growing.
>
> Joe
>
>
> -- 
> Joseph Landman, Ph.D
> Founder and CEO
> Scalable Informatics LLC,
> email: landman at scalableinformatics.com
> web  : http://www.scalableinformatics.com
>        http://jackrabbit.scalableinformatics.com
> phone: +1 734 786 8423 x121
> fax  : +1 866 888 3112
> cell : +1 734 612 4615
> _______________________________________________
> Beowulf mailing list, Beowulf at beowulf.org
> To change your subscription (digest mode or unsubscribe) visit  
> http://www.beowulf.org/mailman/listinfo/beowulf




More information about the Beowulf mailing list