Re: networked processors and parallelism
Eugene Leitl wrote:
>
> Jeff Fox writes:
>
> > About the only people who understand multiprocessing think
> > that the only thing it is good for is supercomputing FP
> > number crunching. When I would talk about F21 at the
>
> Err, but these are very important applications.
Err, I didn't say they weren't important. What I said was
this was opposite end of spectrum from MISC. If you want
a compeditive state of the art formula one car you cannot
sustitute a freight train locomotive. I am not saying
freight trains are not important! I am just saying
that formula one car will not make a good frieght
train and a frieght train will not make a good formala
one car. It has nothing to do with which is more
imporant. Nothing whatsoever.
What is has to do with is that no one would ever argue
that puling freight on tracks is the goal of a formula
one car is not the same as that of a train. We need
trains and fun fast cars. They don't compete.
On one end of the scale you have supercomputers and
workstation farms designed for doing massive FP array
problems. These machines are by definition 1000x
times more complex than what we are talking about.
I don't say that they are not imporant, I say they
are of a very different nature and designed for
different problems.
> Give Beowulfs some time. Small footprint 64 node machines are already
> being made, it's only a question of time before you'll be able to buy
> a multi-node Beowulf box with stackable CPU boards. There QNX and
> Fiasko, so the kernel footprint will get better, if there's demand for
> it.
I don't disagree. How long will I have to wait for the price
of the PCs nodes to come down to say $5 or for their power
consumption to drop by a factor of 1000? A long wait I
suspect. They are not headed in that direction. They
are competing against bigger trains not smaller cars.
For Integer based problems, that are carefully constructed
what has been shown is not only that something as simple as
a twelve bit integer operation may be more accurate and
faster than a floating point calculation. People often
assume that because everyone else does the calc in FP
that it must be done that way and that the problem
actually requires 1000x as much hardware.
When you find a way to solve the problem with 1000x less
hardware then you have opportunity to use 1000x more nodes
at the same price. Sure 64 PCs in a Beowulfs are powerful
so would a machine with 64,000 F21.
> > parallel processing connection they could never get it
> > at all. These people think 'C' and Unix are toy languages
> > so you can imagine how much they understand of Forth.
>
> Uh, the era of Fortran *is* passing, albeit slowly. They've got
> awfully good Fortran compilers for the Alpha, and there's still lots
> of legacy out there.
Yes, but we have never really been trying to penetrate the
frieght train makrket. For one thing the costs in doing so
are much higher than working with things that cost 1000x less.
> Right now the supercomputer market is dominated by Unix and C.
The majority of scientists spending our tax money are still
doing things in Fortran. We heard about this all the time
at PPC and I got the same impression at a NASA seminar on
supercomuting techniques. PC and Unix folks might think
that they dominate supercomputing but they are not the
folks with the big machines or big budgets.
Anyway that is the point we are making. Unix and C and mainframe
like PCs are edging in on the super computing market. They
carry a huge amount of baggage to try to take the high ground.
To carry that baggage you need a frieght train. The whole
idea is in almost the opposite direction as MISC. I don't
say it isn't important. What I say is that I don't see
them taking the ground that we targeted. No one else has
targeted the territory we are aimed at.
> > They think you need a lot more for what they do than you
> > need for C or Unix. They really can't begin to grasp
> > the concept of something with 100 times less than C and
> > Unix.
>
> Jeff, F21 is very much useless for scientific computing, apart from
> very very few exceptions.
I think it it as not "very much useless." ;-) One of the most
demanding and highly scientific computing problems is VLSI CAD
design. We have shown we can do it. We can do it faster, and
more accurately with roughtly 1000x less hardware and 1000x less
code. Just one example of a "scientific" problem.
Science does not equate to FP and gigantic array problems. There
is a lot of room in scientific thinking. Although I realize
that most scientists are not computer scientists and just use
the same computing methods as everyone else. That doesn't mean
that we haven't demonstarted an alternative approach.
> Moroever, the price of a shelved Linux box, bought in quantities, can
> be surprisingly low. Systems on MISC are not exactly cheap, nor easily
> available.
I will accept that when I see a $5 PC that will fit inside of a mouse
form factor.
> I think the future of MISC is programmable logic and minimal systems,
> which have a niche for embeddeds, cryptography, etc.
Another niche. Perhaps that niche has a strong future I don't
know.
My take is that all programs encode knowledege and intelligence.
It is a matter of degree. Conventional programming styles
result in very little intelligence per transistor or byte.
Programs that encode a small amount of knowledge become huge
and brittle. We want to make it possible to encode knowledge
into a machine optimized for this purpose so that the
software can be much more intelligent without having to
require 1000x more code than is really needed or 1000x more
transistors than are really needed.
Of course the designs that have 1000x more resources so that
they can handle a wider range of problems including FP and
huge arrays can perform the same computations that we do
but they just can't solve the same problems in the real
world because costing 100x more they can't be used to
attack the same problems.
You don't embbed a cray in your toaster. Sure it could handle
it but does a toaster really need millions of dollars of
computing hardware? I don't think so.
Jeff
