Pentium bug

[David G. Hough at validgh]

I regret that your column "So what's all the fuss about Pentium's flaw?"
this morning on the Business
Section front page of the San Jose Mercury News completely misses
the significance of the Pentium bug.
The real front page headline "Pentium's flaw may have been hard to prevent"
is more to the point in its implications.
 
Andy Grove does not publish USENET apologies except in a REAL CRISIS.
So we have evidence about what Intel really believes, despite their upbeat
press releases.

The fact that the Pentium bug will affect few people directly is NOT
important.    The fact that nobody knows who those few people are IS important.
There is no way to prove that a numerical calculation on a Pentium 
machine has not been adversely affected by the bug except by repeating
the calculation on a different kind of computer such as a 486.   So the speed
of critical calculations is limited to that of the 486; no improvement at
all.    Calling Intel is no help either in the near term; last week a 
colleague was told that 150,000 people were ahead of him in line for
replacements.  

The worst case of the Pentium bug can result in errors of
several dollars in quantities as small as $100,000.    This is large enough
to get accountants if not the IRS interested.

And in general, anybody who is in some kind of legal jeopardy for incorrect
decisions must be extremely concerned.     Would you want to ride in an
airplane which enjoyed design verification on a Pentium system?    The whole
goal of the current generation of commercial aircraft is to get as close to
the envelope as possible in order to optimize demands for fuel economy,
noise, labor costs, carrying capacity, etc.

You are correct to observe that it has always been true that any kind of
computer decision should be confirmed by independent means.   In the most
important cases, that means on a different machine, with a different
instruction set, running a different operating system, with a program written
in a different computer language, implementing a different algorithm; 
all done by
a different investigator on another continent who thinks in a different
natural language.

It is also correct to observe that while computer hardware and software makers
generally include fine print disclaiming the suitability of their products
for any particular purpose, the majority of computer users these days are
unaware of such issues.     The Pentium bug is a great day for the computer
industry, because it brings to the attention of sellers and buyers the great
cost of getting things wrong.    A primary reason sellers get things wrong
is fear of being late to market, a fear justified by the actions of buyers
who are all too willing to buy the latest and greatest before it has been
fully debugged.

As it turns out, the Pentium hardware bug could have been detected by Intel
at any time by running a test program for a couple of minutes implementing
lecture notes by Prof. W. Kahan of the University of California, who as it
happens has served Intel as a consultant from time to time.   Ironically,
that style of test program is designed to detect very small discrepancies
of the order of 10**-15, and was never intended to catch gross 10**-5 
errors like the Pentium bug.   That it did should further 
undermine the confidence of anybody
who believes that persons who encounter the Pentium bug in realistic
applications are "so unlucky they have already been run over by a truck."