Re: 1teraflops cell processor possible?

Robert Myers <[EMAIL PROTECTED]> wrote:
> On Fri, 28 Nov 2003 08:49:08 GMT, [EMAIL PROTECTED] wrote:
>>
>>I would have to disagree.  My PhD was in protein structure prediction
>>and now I work in genomics and statistical genetics.  Revolutionary
>>advances in these fields are going to come from new algorithms, not
>>from running existing algorithms on faster computers.  Faster
>>computers sometimes do help, of course.  But I think throwing money at
>>the hardware side of the problem now is a waste, since we don't really
>>know yet what the "right" algorithms will look like.
>>

> You are entitled to your opinion, and I respectfully disagree with it.

> If you haven't already done so, please read my post of 11/24 in
> response to Patrick Schaaf, and the IBM document that is cited
> therein:

> www.research.ibm.com/journal/sj/402/allen.pdf

> My conclusion:

> RM>The problems they face are too numerous to discuss in a single web
> RM>post, but the bottom line is that they will have to use up months, 
> RM>if not years, of computing time to get results using models that 
> RM>are at best educated guesses.

Well yes I read this.  The IBM paper is not a bad high level overview.
And I agree with this part of your analysis: after all, this is
exactly what researchers in the field do now.  They develop models,
test them using months of computing time, then try to come up with
better models.  I did some of that too.

> RM>If there is a parallel in the history of science, I am not aware of
> RM>it.  The US wants to build the world's biggest computer, IBM wants 
> RM>to build it for them, and both need a problem that justifies such 
> RM>an enormous expenditure of money and talent.  The conclusion that 
> RM>they should reach, that the available computational muscle 
> RM>available to them is not up to the task they have proposed is one 
> RM>they are unwilling to reach.

It is this part where I think you're wrong.  Blue Gene will be used to
do some good science, I'm sure, and it will be able to do some things
that are prohibitively time consuming on other platforms.  I would
also expect that most of its cycles will be spent doing things that in
retrospect will be seen as a waste of time.  We seem to agree it won't
be revolutionary.  You think this argues for revolutionary approaches
on the hardware side.  I think people will eventually become smarter
about formulating the problem.  Progress on the hardware side is
already fast enough and I see no urgent need to throw more resources
at that side of the problem.

-- Dave