Speculations about the future of the computer industry, and computer arithmetic

[David G. Hough on validgh]

Note that none of the following is based on any relevant inside
knowledge since I don't have any: it's all based on what I read in the
papers, in Unigram, and in comp.arch.

Yesterday's announcement of a limited technology alliance between Sun
and HP, in the area of distributed object-oriented computing, plus the
ongoing minor announcements and major rumors involving some combination
of Compaq, DEC, Microsoft, MIPS, SCO, SGI, lead me to speculate on how
it all should fall out eventually.  Not that it WILL necessarily fall
out this way: there are a lot of egos in the way of economic
efficiency.

If you believe that there will be some kind of potent business
combination based around the MIPS instruction set architecture, and
that IBM is always a pretty potent business combination in itself, and
that the computer industry has twice as many people as it needs - an
observation Scott McNealy makes based on revenue per employee of
rapidly-growing and slowly-growing computer outfits - then you have to
wonder how the rest of the industry is going to survive and prosper.
Apple, HP, IBM, Intel, and Sun all believe they have the critical mass
to survive and thrive, but I have to wonder about the form survival
will take.

It would seem that even Intel can't afford to pursue all its other
lines of business, in particular the i860 and i960, and still put the
resources into the x86 necessary to keep its implementations
competitive with RISC architectures.  Apple allegedly has an 88K
Macintosh in its future, presumably for performance reasons.   HP is
finally putting a lot of energy into putting PA-RISC competitively on
the desktop.

That seems like a lot of effort for RISC architectures that don't have
a large installed base and will have an uphill battle for mindshare
against SPARC and R6000 and a MIPS-ISA combination.   Since the
technical advantages of the various RISC instruction set architectures,
relative to one another, are probably less significant than the amount
of effort being expended on implementations, it suggests that fewer
instruction set architectures might have a better chance for individual
survival.

Ultimately only IBM or Japan Inc. has got the kind of money Intel
needs; since most of Japan Inc. is pretty well committed to one or
another RISC architecture and does not share IBM's need to preserve the
x86, it strikes me as eminently logical for IBM and Intel to resume
their previous joint adventure, spinning off the distractions like the
i860 and i960 that are getting in their way and focusing on enhancing
x86 and R6000.  Particularly if Intel and Microsoft have differences of
opinion about the MIPS combination, IBM's rather large software
establishment might look attractive to Intel.

For added weight in competition against IBM-Intel and MIPS-ISA, it
seems like Apple, HP, and Sun would do well to collaborate to advance
SPARC.  Apple contributes its Macintosh interface, consumer channel,
and applications expertise.  HP contributes its semiconductor,
calculator, and non-computer businesses whose differing market
characteristics would have a stabilizing effect on the financials of a
combined operation.  Sun contributes its SPARC installed base; it's
interesting to speculate whether SPARC would become more open or more
proprietary in such a combination; if the latter, what would happen to
the SPARC clones?  Apple, HP, and Sun should spin off their technology
specific to Intel, Motorola, PA-RISC, and older HP architectures to a
new company which would serve its various installed bases but wouldn't
be likely to grow as rapidly as the whole market.

What about Cray, Europe Inc., Japan Inc., ...?  I think Cray would have
to fit in with one of the three collaborations.  If the potential
members of Europe Inc. could get together and agree on a home-grown
technology such as Acorn or Inmos perhaps they might be able to impose
it on Europe, which is a big and rapidly-growing market.   Most of
Japan Inc. is already committed to some American RISC architecture for
the immediate future although there's plenty of speculation about
whether a competitive home-grown technology is possible or probable and
how the member of Japan Inc. would react.

What does all this have to do with computer arithmetic?  Boiling
everything down to four implementations of IEEE arithmetic, three of
them RISCy, would simplify matters quite a bit in terms of targeting
mathematical software production.  In fact I'd claim that you could
ignore the Intel x86 architecture for mathematical software purposes;
the installed base that you would be interested in capturing doesn't
depend on any IEEE features or even having floating-point hardware
present, and the principal proponent of the x86 would have another
architecture intended for real computational applications.

So if you have three extremely similar RISC IEEE architectures, none
currently implemented with data types more precise than double
precision, there's a good chance that floating-point architecture would
just freeze there for quite a while.  This would probably mean that any
kind of genuine innovation such as efficient support of interval
arithmetic might be further away than ever.

Remember I'm just speculating and have no information that the
foregoing or anything like it is or is not being actively discussed by
Sun or by anybody else.  In a way I have a lot of confidence that human
ego will continue to insure a variety of competitive architectures, and
I have even more confidence that marketing science will continue to
find major points of difference among them even when technical analysis
cannot.