Re: 1teraflops cell processor possible?

Robert Myers  <[EMAIL PROTECTED]> wrote:
>[EMAIL PROTECTED] (George William Herbert) wrote:
><snip>
>>If there is money to be made now, and science and academic
>>careers to be advanced now, using today's technology and
>>algorithms, why on earth should people delay working
>>on the problems with today's solutions just because we
>>know that eventually, hopefully, we'll be able to do it
>>so much better and faster?  
>>
>Sometime within the last year, Andy Glew raised the question on
>comp.arch of what was holding us back in the area of parallel
>computation.
>
>*No one* mentioned a lack of processing power.  The consensus was: a
>lack of solid, workable *software* tools.  That on a forum dedicated
>to hardware.

You're running right past the point.

Large scale generalized parallel computation is a known hard
problem which is not solved.  Known, acknowledged, and it is in
fact receiving a fair amount of R&D effort in the 'computer science'
corner of things.  Plenty of people thinking about it.

Hard problem.  Not moving very fast.

Shrug.  People working on it.  Bright people.  Not my thing;
definitely valuable but it either happens or it doesn't.
Probably will.  Hopefully sooner rather than later.

In the meantime: existing MPP and cluster systems exist,
and existing MPP and cluster software exists, and there are
applications out there using those systems and that software.
Applications which are doing important things for science,
making some people money, helping us avoid nuclear testing.
In many cases these are applications such as atmospheric
modeling or finite element analysis where the problem is
inherently very attackable by a MPP system, without
much software/algorithms pain and suffering.

Many of those applications run much better on bigger MPP boxes.

These systems are not being built and then sitting idle.
They're being used, in production, by real researchers,
real scientists, real engineers.  Making real money and
producing real papers.  And they'll make more real money
and more papers with more CPU power / more nodes / bigger
MPP boxes in a lot of cases.

Some of those applications that scale up nicely include
brute-forcing some of the things that we would really
prefer to find better parallel algorithms for.
But some of them do scale up as you increase the force
applied in a brute force approach.

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

If you build a streaming architecture and can't program it,
the energy performance is meaningless.

If we could program it, we could program a lot of other things
more efficiently, the MPP systems would jump in performance,
and money would come available rapidly to build streaming
architectures if they were demonstrated to then be superior
using those better algorithms.  But if their entire viability
is predicated on software and algorithm and methods developments
which have not happened yet, there is ZERO reason to start
working on hardware now.

>I'm glad that they are possessed of such a profound and instantaneous
>grasp of all that is possible in computation.  I'm a little slower,
>and I think most of the rest of the world is, too, and I'd like to see
>a little more money go into questions like that and a lot less into
>high Linpack scores.

Those Linpack scores are not going into systems going
onto people's shelves to make nice blinkylight boxes.
For the most part, people are using them for real
hard problems, and they're solving things with them.

This is where you fail.

Those systems and problems do not encompass the whole
range we'd like to be working on.  True.

But they're working on real, important problems,
and making real, important progress.  And even though
it's inefficient and unoptimized, it works, and is the
best path forwards that we have available for the near
term for those problems.

You keep denying the existence of that subset of the
total problemspace which they're being used for.
You need to look around some more.  They.  Are.  In.  Use.
Building machines for users who use them is a
valuable activity.  And it is not an activity that
should be put on hold pending speculative developments
on the theoretical side of computer science in making
parallelism work much better.


-george william herbert
[EMAIL PROTECTED]