ALAN@MIT-MC@sri-unix.UUCP (08/29/83)
From: Alan Bawden <ALAN @ MIT-MC>
Date: Thu 18 Aug 83 13:46:13-PDT
From: David Rogers <DRogers at SUMEX-AIM.ARPA>
The closest hardware I am aware of is called the Connection
Machine, and is begin developed at MIT by Alan Bawden, Dave
Christman, and Danny Hillis ...
also Tom Knight, David Chapman, Brewster Kahle, Carl Feynman, Cliff
Lasser, and Jon Taft. Danny Hillis provided the original ideas, his
is the name to remember.
The project involves building a model with about 2^10 processors.
The prototype Connection Machine was designed to have 2^20 processors,
although 2^10 might be a good size to actually build to test the idea.
One way to arrive at a superficial understanding of the Connection
Machine would be to imagine augmenting a NETL machine with the ability
to pass addresses (or "pointers") as well as simple markers. This
permits the Connection Machine to perform even more complex pattern
matching on semantic-network-like databases. The detection of any
kind of cycle (find all people who are employed by their own fathers),
is the canonical example of something this extension allows.
But thats only one way to program a Connection Machine. In fact, the
thing seems to be a rather general parallel processor.
MIT AI Memo #646, "The Connection Machine" by Danny Hillis, is still a
perfectly good reference for the general principles behind the
Connection Machine, despite the fact that the hardware design has
changed a bit since it was written. (The memo is currently being
revised.)