Path: utzoo!utgpu!jarvis.csri.toronto.edu!rutgers!apple!bionet!ames!pasteur!ucbvax!A.ISI.EDU!TKD
From: TKD@A.ISI.EDU (Thomas Donaldson)
Newsgroups: comp.sys.transputer
Subject: where are the workstations?
Message-ID: <12571713354.27.TKD@A.ISI.EDU>
Date: 7 Mar 90 06:44:00 GMT
Sender: daemon@ucbvax.BERKELEY.EDU
Organization: The Internet
Lines: 42


It's way past my bedtime but I felt I should say a bit on the question:
> where are the workstations?

I don't think this is a question about transputers at all, and it is an 
acute question still. Where are the PARALLEL workstations?

Yes, I know that some machines use parallelism, but always in ways built
into the hardware, for graphics or other purposes. And there are some which
aim to be parallel,like the Stardent machine (which came from the marriage
of two clearly unsuccessful attempts at parallel workstations).

Perhpas I am biased by my background (not hardware or computer science as 
such, but numerical analysis). But it seems to me that one MAJOR reason 
hasn't been mentioned at all. The understanding of the best algorithms to 
use on such machines, for some significant large areas people want 
workstations, just hasn't advanced enough to easily support the 
workstation aim. I have my ow
I have my own ideas about what to do about that, but that would be a
completely different comment (yes, as soon as I can I'll get onto it ...
I don't want to be a teaser!). But for instance: FEA methods are in a
parlous state, with only a few algorithms known to use the parallelism at all
well, and those on specific classes of problems. Even for dense matrix mathj
tremendous areas are simply unexplored.

The issue isn't whether or not these things can be ported to a transputer
system. It's easy to PORT. What you need to do is to have something which
takes significant advantage of the parallelism ... otherwise an astute 
customer will notice that he can do it just as fast on a cheaper, nonparallel
machine. And just because yhou have N processors you're not going to find
it simple to run N times faster.

Yes, I know about Rockfield Software. I'd be very interested to discuss nitty-
gritty with them, not their salesmen but those responsible for the 
software design.

I'm not saying these problems are impossible to solve, but they are hard, 
and any investor or worker in the area should keep this in mind and be 
prepared for setbacks along the way.

			Thomas Donaldson
-------