Re: COMP: A DPA of VPU's to replace CPU's?

Carl Feynman (
Fri, 22 Aug 1997 12:09:23 -0400

At 01:54 PM 8/21/97 -0400, wrote:

>Now I need to know:
>Has it been done before?

Yes. You have described the Connection Machine, models CM-1 and CM-2,
manufactured by the Thinking Machines Corporation of Cambridge Massachusetts
from 1984 to 1990, and designed by yours truly and some other people. Also
the MasPar NCube and the Meiko Computing Surface, and a bunch of others, all
built during the same period. They sold adequately, but mostly to the
military-industrial complex, so when the cold war ended, so did the sales.

>And will it work?

Well, it seemed like a good idea at the time, but it turned out that the
right size for a processor is about 32 bits wide, not 1 or 4 bits, and it
turned out that the manufacturing techniques for RAM and processor were
different enough that you're better off putting them on seperate chips.
Then, given that you had designed a 32-bit processor on a chip, you might as
well give up on the central instruction broadcast trick and use a separate
instruction fetch engine on every chip, because it buys lots of flexibility
at little cost. That's close enough to a regular microprocessor that you're
better off just buying them, and adding special-purpose logic for
interprocessor communication. The upshot is the design of the CM-5 and
every other post-1990 large parallel computer: hundreds or thousands of
standard one-chip CPUs, each with their own memory, and some special-purpose
network hardware conecting them.