[geeks] anyone know about this? 72-core, 48GB computer?

[gsm at mendelson.com]

On Thu, Oct 01, 2009 at 02:38:55PM -0400, Joshua Boyd wrote:

>I poo-poo'ed the SGI for almost literally being a fancy case for off the
>shelf boards, and failing to bring anything new to the party, unlike the
>cray system, which makes infiniband standard and offers a nice mix of
>graphics and tesla options to go with the computer blades.

Beowoulf in a box? Not intertesting to you, but may be interesting to
someone who wants to buy a cluster off the shelf. It's a shame that SGI
is reduced to doing so. 


>However, in addition to liking MIPS, this system uses a proprietary
>interconnect which could be interesting, if for no other reason than the
>somewhat odd connection graph (Kautz) that appears to be made of
>interconnected rings.

What happened to that? Now that the company folded, did the technology end
up being bought? Will it sit in patent hell until they expire or no one
pays the maintainance fees? 


>
>It is also possible that there are nice instruction set properties
>compared to Atoms.

I don't know what's in the dual core ATOM's?


>Finding people skilled at writing HPC software is hard enough that
>writing for Intel versus MIPS versus PPC versus something new is little
>difference. 

You must know different people than I do. Most of the people I know learned
HPC computing on Intel (MOSIX was developed here at Hebrew U), and there
are zillions of beowoulf clusters out there based on PC's. I knew someone 
who put together one based on Alpha PC clones (PC clones with Alpha instead
of Intel chips), but they are all probably long gone.

>Or at least no one was interested in investing the money in advance of
>getting hardware.

I was thinking of venture capital instead of pay-before-delivery sales.

>You would also need to add in a PCI-8x infiniband controller, or
>something else like myrinet, or something new.

Sure, there are lots of them out there.


>
>> Of course, you could just buy a bunch of ASUS EEE motherboards, with a  
>> rack type power supply and a 1000BASE-T hub in a cabinet to make a proof 
>> of
>> concept unit. :-)
>
>You could, but it wouldn't prove much.  It would only be interesting if
>you used a fancier interconnect.

It could prove you could do it, and get venture capital. The fancy interconnect
would be on the first run of boards, which would be made to seed the market
and a second round of funding obtained. 

Basicly what these people failed to do. Of course we have no idea why, they
may have found techology problems, lost their source of processor chips, run 
up against patent issues, fought among themselves, or any number of reasons,
probably most of which we will never know.

Geoff.

-- 
Geoffrey S. Mendelson, Jerusalem, Israel gsm at mendelson.com  N3OWJ/4X1GM