Re: 1teraflops cell processor possible?

> He posted back that parallel Matlab was a subject of research, not a 
> tool for working scientists, that the program he was using represented
> man-years of development effort, and that he had neither the skill nor
> the time to retool it for use on even a two-processor machine.  

There is the "wrong turn" of spending too much time on your tools, and
not getting your real work done. There is the equally wrong turn of not
spending enough time on your tools, and this guy belongs in this category.
As you say, he couldn't even be bothered to search for possible parallel
implementations of what he needed. If he was working on a PhD, he should
be changing advisor, as his is not doing the job properly.

> If the DOE, or anyone else, were providing a realistic level of
> funding for basic research in parallel computation, I would not be so
> offended at their throwing however many million at just another big
> machine.

I believe they did, but not very much came out of it. It takes very
strong leadership and high expertise to get the correct balance in such
a program of research - be experimental, follow new ideas, but don't 
fall for fads. See David de Nucci's reported experience in finding 
funding. And such a program needs to be backed up with sufficient hardware
support.

> If you've got a neato idea, and it will make zillions of bucks in
> biotech, go get some venture capital and do it.  If your idea is for
> real, you'll find the money.

In the current economic climate? No way no how.

> For all the bilge and bother of this thread, I'm left with a very
> basic question: can you get the energy performance out of a classical
> architecture that you can get out of a streaming architecture.  

Define "classical". 

> Two posters have stated without proof that they *know* that a streaming
> architecture won't beat a classical architecture on realistic code.

>From your description, "streaming" basically means re-using operands as
much as possible, by passing the result(s) of one FU to the next FU with
as little intermediate storage as possible. What the "two posters" told
you was that current microprocessors already do that if the programmer
can make it happen, and the energy cost of reading and writing registers
and L1 is negligible compared to other costs.

So the difference should be made at intermediate and long distances of
communication - L2/L3, off-ship/memory, etc. But then, we argue, BlueGene
is already a strong departure from the model of the "classical" micro-
processor - whatever that may mean in detail. As was the transputer,
for instance, or the ICL DAP. Or, indeed, the MasPar, the CNAPS, and
the venerable CM-1.

And we agree that the real problem is on the software side. But I, at
least, disagree that it's a problem of theoretical research: you need
to get people involved with the new ideas and the skills in models of
parallel computation, _and_ the people developing the applications- but
see above on the readiness of the application side to get involved. 

All too often in computer "science", things have been tried on a toy 
scale and found to be "good" (at least good enough for yet another M.S.
or PhD), but to collapse on first contact with the "real world". With
a background in image processing, I can tell you that this is a rampant
disease. Also see my previous remark about the perceived value of inter-
disciplinary research. You need to change the cultural values of these
areas (!) of science before you'll achieve anything.

        Jan