bruce@godot.UUCP (Bruce Nemnich) (06/24/84)
Wow, what a discussion; I finally caught up.
I totally agree with Doug and Barry: just buy a bunch of lisp machines.
Here at Thinking Machines, our computing resources consist of ten lisp
machines (with 5 more waiting in the wings at our new building) and a
vax 750. The vax is here for communication (mail, news) and document
editing, not programming. Originally, it was intended to be the main
file-server for the lispms, but people tend to like the lispm file
system better. I am the only person who programs on this system, and I
mostly just fix bugs and write convenience utilities to keep things
running smoothly.
The lispm environment is BY FAR the most productive environment I have
ever used, by at least an order of magnitude. There is no amount of
hardware or software available at any price which would make our people
nearly as productive. Incidentally, we find we need about 2/3 of a
lispm (Symbolics 3600 or LMI Lambda) per full-time person to ensure that
machines are available to those who want them, given the variable
schedules that we have.
Several articles ago in this chain, some skeptic implied that it would
be difficult or impossible to run a lisp-like language on a non-von
Neumann machine. We are building a very-fine-grained parallel
architecture called the Connection Machine, which will be totally
controlled by lisp. The Connection Machine was designed exactly because
lots of AI problems which are totally impractical on conventional
machines are extremely parallel in nature (e.g., searching a big
semantic network for a node of information). We wouldn't think of using
any other language.
Von Neumann or not, the key to getting an efficient implementation of
lisp is to get the architecture right. The efficiencies you are
quibbling over are within an order of magnitude; we will get 5 to 7
orders of magnitude on the problems we're building this machine for.
--
--Bruce Nemnich, Thinking Machines Corporation, Waltham, MA
{decvax!cca,ihnp4!mit-eddie,allegra!ias}!godot!bruce, BJN@MIT-MC.ARPA