[rescue] SMP on intel wasteful?

[Dave McGuire]

On June 26, Francisco Javier Mesa-Martinez wrote:
> >   Well, not quite enough said.  Looking at history you will find (and
> > I'm sure you already know this) that vector processing was just about
> > the only way to get really serious cost-is-no-object numeric
> > performance out of a single processor until fairly recently (past 15
> > years or so maybe?).  Since then scalar processors have become much
> > MUCH faster, not meaning to state the obvious.
...
> 
> Note: Just to make clear.. I have nothing against vector machines. Hell I
> even have a CRAY1 manual. Is just that I got the sad feeling that these
> machines are a dying breed... maybe they are more prevalent in other
> fields, but at least in my academic architecture research env they have
> been pretty much written off. Of course now the lines are getting fuzzy
> again, as modern processors include now "vector" (I would not really call
> them that but that is what marketdroids like to call them) units. So at
> least some lessons have been learned :-).

  Well I know that you have a lot of experience in this area, and I know
you're not a bullshit artist...so I'll interpret this to mean that, in
your area of research, vector processing isn't the Right Thing.  And
that's ok.  Please don't interpret my bad mood for argumentativeness
here.

  Though, to further expand upon the point I tried to make above, this
is more what I had intended to get across.  Since, up until maybe
fifteen years ago, vector processing was the best way to get good
numeric performance, lots of applications that weren't necessarily
best addressed in that way were shoehorned into vector machines.  It
is my opinion that the shift you indicated reflects the movement of
those applications to multiprocessor scalar machines (or clusters for
that matter) that are better suited to these problems, and are now
fast enough to address these problems effectively.

> I guess it all comes down to how $%#%ing hard it is to compete against
> micros with their economy of scale. I hate Intel's cludge of an
> architecture.... but one has to be impressed at their fab/process research
> and capabilities.

  Their process technology is damn impressive.  But where can they go
from here?  Itanium is dead, Pentium-4 is the bastard stepchild of
Pentium-III...Pentium-5 perhaps, with more castration and higher clock
rates?  This isn't evolution...it's milking.

         -Dave

-- 
Dave McGuire                  "Needing a calculator indicates that
St. Petersburg, FL              your .emacs file is incomplete." -Joshua Boyd