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

[Mike Meredith]

On Thu, 1 Oct 2009 20:07:10 +0200, gsm at mendelson.com wrote:
> On Thu, Oct 01, 2009 at 01:32:50PM -0400, Joshua Boyd wrote:
> >
> >I don't think that the SC machines are single system image.  I think
> >they run an image for every 6 processors.  I don't know that for

A quick look at the spec sheet indicates that no it's not a single
system image, but it seems to do the management of the nodes in a
sensible way ... enough that it's a little more interesting than a
standard beowulf cluster.

> As much as Intel processors are un-interesting at least they are
> standard and which makes them well tested and easy to find skilled
> programmers for.

How much hpc code is written in assembler anyway ? As long as you have
a decent FORTRAN compiler (and probably C(++) too these days), your
problem is solved. Besides you'll have far more trouble multi-threading
an application to take advantage of 72 cores (or 12x6 cores) than
dealing with an "unusual" instruction set. Besides MIPS cores are
pretty widely used in the embedded market (or used to be). 

> Personally, I'd love to see a quad core ATOM (which I assume is in the
> pipeline, but that's pure speculation) on a little board that plugs
> into a backplane. Then you could do the same thing with processors
> that everyone knows and loves. 

And you have a standard backplane that'll provide a _fast_ interconnect
between nodes ? It's not the number of nodes that's interesting; it's
how fast those nodes are connected (except for a subset of hpc
problems).

-- 
Mike Meredith (http://zonky.org/)
 "Microsoft couldn't get a clue if it were soaked in clue pheromone in
  clue bondage gear on the clue mating grounds during clue mating season
  surrounded by horny clues." -- corrupted from a SDM sig