Transputer documentation required

From: Doug Yowza <yowza_at_yowza.com>
Date: Fri Jul 17 05:57:14 1998

On Fri, 17 Jul 1998, Julian Richardson wrote:

> I was tempted to ask how fast these 8088-based 'cubes were, but I
> suppose it's really hard to describe real-world performance of such
> systems without actually using them...

Not very fast, but Intel refined the idea, and I believe they had the
world's fastest supercomputer at one point. The basic idea was that you
take two elements: a processor and a communication channel, put a bunch of
them together is a hypercube topology (in which each node talks to its N
nearest neighbors in an N-dimensional cube), and then partition your
problem so the processors could work on it in parallel.

So, in theory, a 5-cube with 32 processors could run up to 32 times faster
than a single-CPU computer. In practice, communication overhead between
the nodes can kill performance, and it's not always easy to (re)structure
a problem to take advantage of the parallelism.

I never worked directly with transputers, but I think the topology was
pretty similar, and the operating environment was based on something
called Occam. Somebody else with Occam experience should chime in
here, but it was basically a language and underlying multiprocess OS that
allowed programs to automatically scale to as many processors as were
present.

-- Doug
Received on Fri Jul 17 1998 - 05:57:14 BST