Relay-Version: version B 2.10 5/3/83; site utzoo.UUCP
Path: utzoo!mnetor!seismo!rochester!pt.cs.cmu.edu!spice.cs.cmu.edu!ram
From: ram@spice.cs.cmu.edu (Rob MacLachlan)
Newsgroups: comp.arch
Subject: Re: Connection Machine Argument
Message-ID: <1097@spice.cs.cmu.edu>
Date: Wed, 3-Dec-86 17:19:49 EST
Article-I.D.: spice.1097
Posted: Wed Dec  3 17:19:49 1986
Date-Received: Wed, 3-Dec-86 22:29:11 EST
References: <745@husc6.UUCP>
Distribution: na
Organization: Carnegie-Mellon University, CS/RI
Lines: 22


    The connection machine is not committed to a hypercube architecture.
Routing of messages is done by the hardware below the instruction-set level.
Of course the actual cost of sending a message to a particular address will
depend on the topology of the connection network.  The existing languages
for programming the CM don't allow the underlying network topology to be
seen.

I would say that you have misunderstood the idea behind the CM, at least at
the routing level.  I think a better summary would be to say that there are
O(n) processors connected by an appropriate interconnect with a smart
controller.  Intelligence and flexibility at the routing level is definitely
part of the original concept, although it may not be fully realized in the
current design due to a desire to get something out the door.

There certainly is something to the claim that 1bit processors aren't the
ultimate for everything, but they seem to be well suited to the symbolic
processing tasks that the CM was originally designed for.  For things such
as floating point, it would be quite possible to associate a floating point
processor with each bit-serial processor.

  Rob