Re: 1teraflops cell processor possible?

On Tue, 2 Dec 2003 10:03:09 -0600, "Del  Cecchi" <[EMAIL PROTECTED]>
wrote:

>In just rereading the IBM J. of Research and Development article
>http://www.research.ibm.com/journal/sj/402/allen.html
>
>I was struck by the following paragraph.
>--------------------------------------------------
>More generally, thinking on algorithm and system design is still dominated
>by the invalid perception that compute units are the most expensive resource
>in a computing system. Applications are designed to minimize the number of
>operations required to solve a problem. Systems are designed with enough
>memory and I/O to achieve a high usage of the compute units, even if this
>leads to low memory or disk utilization. In fact, the cost of systems is
>dominated by storage cost, and by the cost of the logic required to move
>data around (caches and buses). If one thinks anew from basic principles,
>the conclusion will often be that algorithms that use more computing but
>less memory, or use more computing but require less communication, coupled
>with systems designed to ensure more effective use of memory and
>communication, even at the expense of lower utilization of the processing
>units, are more cost-effective than conventional algorithms and systems.
>-------------------------------------------------------
>
>This is from 1999.  Is this the paper referred to earlier?
>
>
I referred to the document and provided a link to allen.pdf in a Nov
24 post responding to Patrick Schaaf.  In my post I tried to give both
IBM and the creators of the document their due for being both candid
and thorough.  I was critical of the document for not reaching the
conclusion that I think a presenter not eager to make a sale would
reach; viz, that while we know alot about the problem and we (IBM) are
very competent at designing big machines, the machine we (IBM) are
proposing is not equal to the problem we are proposing.

You should understand that the perspective I have been expressing is
not uniquely directed at Blue Gene/L.  I have already made some very
strongly critical comments about the Space Shuttle.  I was not all
that unhappy when the Superconducting Super Collider was cancelled
because, as I understood it at the time, there was a significant
chance that it would return a null result on the questions that
justified such a large expenditure on basic science: the top quark and
the Higgs boson.

I preferred then and I prefer now to rely more heavily on the natural
particle accelerator that the universe has provided to us and to get
as much instrumentation as possible above the atmosphere so we can
catch the universe in the act as it reveals its secrets to us without
our having to perform unnatural acts on terra firma.

Instead of people asking the question: how much time and money would
we need to do this right, they ask an oddly inverted question: given a
schedule and a budget imposed by considerations that largely have
nothing to do with science, what can we build and what can we
plausibly claim it will do?

The only magic that will suddenly appear at 50Teraflops is that Japan
will become no. 2 and we will become no. 1 again.  That's not a good
way of managing the finite resources available to science.

To get back to the IBM document, I would have been very impressed if
an integrator of big systems had followed the paragraph you quoted
with another paragraph that said: "Since we know almost nothing about
how to construct algorithms that achieve what we have identified as
desirable goals, we therefore know almost nothing about how to
accommodate such algorithms with hardware.  Before investing a large
sum in a small number of large systems, we should spend x years and y
dollars working with a larger number of prototype systems so we can
get a better handle on what kind of large system we really should be
building."

RM