Re: 1teraflops cell processor possible?

Robert Myers  <[EMAIL PROTECTED]> wrote:
>[...]
>Messianic or not, humble or not, I'm standing by my position: spending
>money on monsters like Blue Gene is a waste.  It won't make any real
>progress on the problem it purports to be aiming at (I'm repeating
>myself), and it would be *much* better to have tossed a few of those
>boxes out to universities and let them figure out how to cope with a
>few thousand of what Rupert refers to as mice, rather than putting
>65000+ of them together to see what comes of it, other than a
>meaningless Linpack score.

I don't get this.

You're pissed off because the notional declassified cover use
is inefficient, not unproductive but inefficient?

One, there are clear national security reasons to be building
boxes like this, as has been repeatedly pointed out.
Sufficiently accurate simulations of nuclear weapons to
avoid the need for testing, simulations that can include
defects and aging effects and the like, are a very important
national priority.

Two, boxes like this can do other types of work than protein
folding, in terms of FEM and fluid flows and weather analysis
and all sorts of other useful and neat things.  Bigger boxes
can do it more accurately, or faster, or both.

Protein stuff was not the only reason to build them.

Protein stuff is funding, both commercially and in research/academic
environments, some large computing facilities.  The people doing
that work are finding computational proteomics to be a field that
is extremely valuable to advancing theoretical research and commerce.
They are finding it to be an enabling technology even though it's
stuck with using obviously inefficient algorithms at this time.

Systems like Blue Gene will work just as well at brute-forcing
the protein folding and related problems as smaller systems are
at this time.  You seem to be offended that we're attacking the
problem by brute force rather than finesse.  That's fine,
as far as it goes.  Finesse is always useful.  But finesse requires
someone having finessed the problem so we can do it the better
way already.

But your value judgement of the low value of brute-forcing the problem
is not matched by the demonstrated academic, research, and commercial
user base's experience.  I repeat: Experience.  They are not buying
systems to have labs full of cool blinkie lights.  They're doing their
work with it, producing papers and products.

I was involved in building and supporting one commercial
computational biology project, and did proposals for two
others in the hundreds of nodes range, not world-class by
current standards but large.  A noticable fraction of
the workdays of the researchers are being spent waiting
for computational results to come back.  While spending
tens or hundreds of millions of dollars didn't make sense
for any of these customers, there are projects going on in
the field that will justify the use of systems of the
scale of Blue Gene.

That we theoretically should do much better eventually
once we figure out how to finesse the algorithms is of
no import to the calculation of current value of doing
current computational projects in those fields.  If it is
worth doing now, with today's technology and algorithms,
then it will be done.  And it's being done.

If there is money to be made now, and science and academic
careers to be advanced now, using today's technology and
algorithms, why on earth should people delay working
on the problems with today's solutions just because we
know that eventually, hopefully, we'll be able to do it
so much better and faster?  


-george william herbert
[EMAIL PROTECTED]