Re: 1teraflops cell processor possible?

On 30 Nov 2003 17:24:43 -0800, [EMAIL PROTECTED] (George William
Herbert) wrote:

<snip>
>
>I think the problem I see with your attitude on this issue is
>that the truly fundamentally hard problems seem to be to some
>large degree to be algorithms and software rather than hardware.
>
I've been around the mulberry bush at least once on this subject on
comp.arch.  It is almost a religious belief with me that thinking
about hardware and software separately is the road to ruination.  Even
if it's only conceptual hardware, as in Turing's amazing little
machine, you'd better have something concrete in mind. 

>If there were a software pull, someone would push hardware
>in that direction.

The only example that comes to mind is LISP machines, which (and
someone will correct me if I'm wrong) I believe came to nothing.

Other than that, hardware has pulled software, at least in my little
universe of computational physics:

Vector algorithms.
Stream processing.
Cache blocking.
Instruction chaining.

Other than that, it's been a matter of trying to figure out how to get
hardware to do things that it doesn't do very well naturally, like
pointer chasing.

In the case of the current tired horse I've been flogging, it's how
many problems can you cast into a streaming format and what kind of
payoff is there for doing it.

>There is good reason to think that from
>a technical perspective, developing the sorts of software to
>do those problems better would be good, but it's proven rather
>hard in practice, including in PhD thesies and random researchers
>reaching out in freethinking directions.  Even if it had to be
>hand-done to some degree, if it was doable for some of the problems
>there would be money right there for doing it.
>
I do not think that either Church or Turing would be pleased to see
the mess that has been made of their elegant formulations.  I don't
think it would hurt at all to invest in some mathematical
street-cleaning.  If enough of that kind of work were funded, another
Church or Turing would stumble along soon enough.

>Throwing money at hardware problems to run the known algorithms
>in larger parallelism works, to a predictable degree.

and unless you can get a qualitatively better result, you are just
throwing money away.

>To make the
>great leaps, we need some great leaps in concept and algorithm
>which have as of yet at least largely eluded humankind as a whole.
>That is not the sort of problem that you can solve with money.

No, it's better to let starving mathematicians do it on their own so
that lesser figures with swollen egos can come along later and get on
the cover of Time magazine.

>It's waiting the right bright person and the right supporting
>developments for them to make the right insights.
>
And it is somehow better to throw huge sums at arbitrarily large
machines than it is to fund smaller scale but more risky research?

>The question is; given the value of some of those problems,
>is it worthwhile for society to spend money on the ugly hack
>way we do it now being scaled up now, or should we just not bother
>and wait for the great leap that may come?
>
Blue Gene was a knee jerk, and pretty much wasted, response to an
external stimulus.  Somebody needs to be embarrassed.

If you can't actually *do* the real problem (e.g. protein folding),
spend your time on paper studies and wait for the hardware to catch
up, which it will.

>The government wants to do some of those things now, and commercial
>companies are making money off doing some of those things now,
>so I am guessing that the value of those problems is worth the
>incremental improvements we can make now.
>
The US is saving itself the embarrassment of being No. 2 in an
important technology as a direct result of being letting the PC market
fund processor research for over a decade.

>Banging your head on the wall and demanding that the next
>Einstein, Feynman, Hawking pop up and solve 'the problem' is
>not a reasonable  strategic plan 8-)
>
I don't bang my head against the wall, and I'm not a particular
admirer of any of the three figures you named.

Sooner or later, Eugene Miya will reappear and remind me that, after
all, there are some things that just can't be done, and I will again
respond that, if you are persistent, you may not get an exact answer,
but you can come as close as your patience and persistence permit.

RM