Relay-Version: version B 2.10 5/3/83; site utzoo.UUCP
Path: utzoo!linus!decvax!cca!ima!haddock!johnl
From: johnl@haddock.UUCP
Newsgroups: net.arch
Subject: Re: Re: 16k vs 68k vs 432 - (nf)
Message-ID: <21@haddock.UUCP>
Date: Sat, 31-Dec-83 23:35:30 EST
Article-I.D.: haddock.21
Posted: Sat Dec 31 23:35:30 1983
Date-Received: Sun, 1-Jan-84 05:26:17 EST
Lines: 45

#R:utzoo:-342700:haddock:9500006:000:2392
haddock!johnl    Dec 31 13:05:00 1983

Here's a sidelight on the 68k vs. ns16k vs. 80?86 arguments.  I've lately
been looking at floating point chips for microprocessor systems.

The Intel 8087 is numerically wonderful.  It does everything that the
IEEE standard asks.  It's not terribly fast, but the 80287 is a lot
better.  The problem is that the only way you can use one is in tandem
with an 8086 or 8088 (or 286 in the case of the 287.)  It's designed to
be wired in parallel with the other chip, and looks for all sorts of
strange stuff on the bus that is practically impossible to simulate if
your other chip happens not to be Intel's.

The Motorola 68881 is also numerically good, reasonably fast, and allows
practical interface.  It's clearly the best chip for designs not based on
Intel CPU chips.  The only problem is that they're not available.  (If
somebody can prove me wrong on this point, I'd love to hear about it.)

So we're left with the National 16081.  It's fast and easy to interface
to, but numerically it's bad news.  They only implemented the easy part
of the IEEE standard, and left the rest of it up to software.  This means
that if you care about your numerical results, you have to do a lot of
software fixing up, which decreases real performance considerably.
There's no provision for extended precision intermediate results, so if
you want them, you have to simulate them.  Square root and trancendental
functions are up to you, unlike the other two chips which implement them
in microcode.

I few years ago I was at a SIGMICRO meeting in Orlando where an
interesting (and lamentably poorly attended) session was given by the
head of the 8087 design team and two Israelis who were on the 16081 team.
One got the distinct impression that the National chip was designed by
people who didn't understand floating point arithmetic very well, so they
took shortcuts that seemed reasonable to them but make life difficult for
people who have to use the chip.  The Intel guy, on the other hand,
clearly knew his topic top to bottom.  He pointed out that once you look
at realistic floating point code which uses the transcendental functions
and uses the full IEEE standard, the nominal speed advantage of the
National chip is less useful than it seems.

Nonetheless, due to its availability and its relative ease of interface,
the National chip is currently the one of choice.  Sigh.

John Levine, ima!johnl