COMMENT ⊗   VALID 00002 PAGES
C REC  PAGE   DESCRIPTION
C00001 00001
C00002 00002	denelc[w84,jmc]		Why DARPA should buy us a Denelcor machine
C00005 ENDMK
C⊗;
denelc[w84,jmc]		Why DARPA should buy us a Denelcor machine

	It is quite evident that very high speed computing will
require greatly increased parallelism.  However, none of the schemes
for implementing parallelism have been shown to allow full use
of the parallel computing capability on single problems.  Therefore,
there is a considerable risk involved in building a large parallel
machine dedicated to any of the present parallel architectures.
This risk can be reduced by getting programming experience as promptly
as possible.

	One of the proposed parallel architectures
involves many processors, each with independent control, looking at
a large memory.  This is the one favored by John McCarthy and Richard
Gabriel.  In fact we have developed a Lisp oriented toward this
architecture.  The Denelcor machine is the first commercially
available machine with such an architecture, and it is important
to explore its usefulness for the kind of programming done in AI.
The results of this exploration can then be fed into more ambitious
machine design projects.  It will also be important to determine
whether the dataflow features incorporated in the Denelcor, i.e.
the FULL and EMPTY bits are worth their cost.  According to people
at Los Alamos, the Denelcor architecture has proved suitable for many of the
kinds of computation they do.

	A secondary goal is to develop a modern time-shared operating
system for such a machine.

	Because the Denelcor uses a circular pipeline architecture
with one actual processor pretending to be many, a single processor
Denelcor will suffice.  On the other hand, AI work will require
at least two million words of memory, and this is rather expensive
on the Denelcor.