Re: 1teraflops cell processor possible?

On Fri, 28 Nov 2003 00:56:27 -0500, "Bill Todd"
<[EMAIL PROTECTED]> wrote:

>
>"Robert Myers" <[EMAIL PROTECTED]> wrote in message
>news:[EMAIL PROTECTED]
>> On Thu, 27 Nov 2003 17:02:07 -0500, "Bill Todd"
>> <[EMAIL PROTECTED]> wrote:
>
>...
>
>  Unless I greatly over-estimate the expense of
>> >creating large, complex, high-performance chips (and the recent
>experiences
>> >with Itanic would seem to suggest otherwise), doing so for a small,
>> >special-purpose market such as the one described seems to make no sense
>at
>> >all.
>>
>> I certainly don't think you are, but I still don't believe we should
>> be building supercomputers with COTS proceesors.
>
>How can you possibly justify such a statement without performing the kind of
>cost/performance analysis I made at least a very rough stab at?  It would
>seem obvious that there is at least a *possible* range of costs implicit in
>creating new architectures to support supercomputing for which using COTS
>processors instead is eminently more sensible, because while they may offer
>only a small percentage of the efficiency of a custom implementation they
>may do so at sufficiently low cost to make them the better choice.
>
C'mon, Bill, if I have the nerve to call a program that got by high
level review panels, the Congress, and the President, and that the
American people stood by and waved their flags for "just plain
stupid," I'm obviously the sort of person who thinks he can get away
with saying anything. ;-).

Some sort of by-play has been going on between the powers that be and
companies like IBM:

Powers-that-be: "We need faster processors."
Companies like IBM: "We can build them.  Just tell us how to pay for
them."

Powers-that-be: "You know.  The same way you pay for everything else
you design and build."
Companies like IBM: "We build processors for which we can find lots of
buyers.  The kinds of processors you want us to build won't have lots
of buyers."

Powers-that-be: "Well we need those processors.  How do we get them
built?"
Companies like IBM: "Get out your checkbook."

>The only areas that come to mind where it might be worth, say, a 50x
>increase or more in cost to obtain a 10x improvement in performance are
>time-critical computations such as weather forecasting (and my vague
>impression is that this particular activity may be at least somewhat
>amenable to simply increasing the number of processors you throw at it) and
>real-time automated battle analysis.  In other areas, I suspect (with very
>little acquaintance with HPTC to go on) that very often either you can throw
>more (inefficient) general-purpose processors at the problem or you can be
>just a bit more patient (because your time just isn't worth as much as the
>hardware would cost - at least if you manage that time sensibly and divert
>yourself to other useful work while waiting for the result).
>
I was just about to say, "You really need to read...," but I already
pulled that stunt in this thread.

There are forty million people worldwide infected with HIV.  Five
million became infected in the last year according to the Washington
Post.

Computers won't necessarily find a cure or a vaccine for AIDS, but
computers could revolutionize molecular biology the way they
revolutionized aerodynamics.  We are just at the point of being able
to contemplate the undertaking, but we need much more powerful
computers.  We need to start trying harder, and we need to stop being
cheap about it.

>As far as global use of limited resources, I tend to sympathize a lot more
>with scientists who can't obtain the hardware required to get new
>information at all than with those who are merely anxious to get it faster.
>And, for that matter, with many other non-scientific funding needs.  So it
>seems just fine to me to let the free market decide whether you should get
>the hardware you're asking for.
>

Having brought up the worldwide AIDS epidemic, I'm not going to make
the obvious retort, but it begins with "Well, then, why don't we just
let the free market decide...", and I think you know how it would end.

RM