Path: utzoo!attcan!utgpu!jarvis.csri.toronto.edu!rutgers!tut.cis.ohio-state.edu!cica!gatech!hubcap!landman
From: landman@Sun.COM (Howard A. Landman)
Newsgroups: comp.parallel
Subject: Re: Scaled-down CM ?
Message-ID: <5874@hubcap.clemson.edu>
Date: 29 Jun 89 12:14:13 GMT
Sender: fpst@hubcap.clemson.edu
Lines: 32
Approved: parallel@hubcap.clemson.edu

In article <5847@hubcap.clemson.edu> eho@bogey.princeton.edu (Eric Y.W. Ho) writes:
>Does anyone at TMC considered building a somewhat scaled-down CM box -- like
>putting just several hundreds or just a few thousands cells on say a big wafer
>or several large chips so that it can be hooked to one's desktop (either as a
>separate box next to the desktop or actually inside the desktop -- that would
>be really nice) so that people with a limited budget that can take advantage
>of this architecture (& don't want to learn another entirely new enviornment
>-- it is nice that CM can adapt easily to the host's environment) but not
>necessary need so many cells can go out and get one of this box (either for
>personal use or to be shared by a small group of half a dozen or so people)
>and start doing some useful things with it.

Whew!  That's an awfully long sentence!  Anyway, I brought this same issue
up a few months ago.  Here's one problem: the cost of a CM can be divided
into 4 parts.  (1) Host (2) Host interface board (3) Sequencer (4) "Cells",
to use Eric's term.  The host has to be sufficiently powerful to not be the
bottleneck; a small Sun 4 or a VAX 8600 might be O.K.  Both CPU power and
I/O capacity are important.  Either way you're talking at least $10K to $X00K.
The interface board and sequencer are the biggest cost problem - they cost
the same for one processor or 16K.  Let's say another $10K to $100K.  Finally
there are the processor cells.  A good number might be 256.  But each CM
processor is fairly weak (maybe 0.1 MIP) by itself - it's only in large
numbers that they start becoming powerful.  A 256 processor CM would have
a peak, maximally utilized power of about 25 MIPS.  Not very exciting for
a couple hundred $K.

If you really want to play with what it's like to program a CM, it would
probably be better to just write StarLisp and Fortran 8X and C* compilers
for conventional machines, and buy a fast deskside to run them on.

	Howard A. Landman
	landman@sun.com