Kort Explores conundrums of the human condition

On page D-1 of their 12/21/94 edition, the Los Angeles Times ran an article publicizing a flaw in the Intel Pentium chip that could cause calculation errors under some conditions. The bug appears to be caused by 5 missing entries in a look-up table implemented in a PLA (programmable logic array). The result is the loss of precision in certain floating point calculations involving large numbers.

One has to wonder at the ferocity of the media attack on Intel - especially since the most vocal detractors in the computer industry have a vested interest in products that directly compete with Intel. It's hardly news that CPU chips have had flaws. According to some experts nearly all of the computers ever built have equally serious microprocessor flaws. "Serious" is obviously a relative term.

In fact, prior to the formulation of the IEEE standard floating-point formats, the numeric results generated by different brands of computers rarely agreed. These hardware computational errors were usually (often reluctantly) compensated for in software. Lots of computer owners never make use of their floating point unit anyway. Since I deal primarily in symbolic logic and predicate calculus, and have limited need for heavy duty number crunching, none of my software products (or my users) would ever be effected by the Pentium flaw.

At first Intel didn't take the problem seriously. It had calculated that the flaw would only effect the average user once in 27,000 years, and had fixed the problem in a subsequent stepping anyway. They started personally contacting those users they expected to have problems to quietly replace their chips. However, IBM calculated the average user would experience the bug once in every 28 days, and very publicly announced it was suspending all shipments of Pentium products until the chip was replaced. The tide of public opinion forced Intel to offer to replace any Pentium chip with the flaw free of charge.

Unfortunately, the LA Times article is as badly flawed in its calculations as the Pentium problem it targets. According to the article, you can test your computer for the Pentium flaw by calculating the following formula: (4,195,835*256)/(3,145,727/256)

If your computer has a flawed Pentium CPU, the result will be 87,407.93. According to the article, the correct result is 87,413.1. I don't have a Pentium myself - my "computer" is composed of 2 networked 486DX's (33 & 50Mhz) running OS/2 (2.1 & Warp) on the theory that 2 sort of fast computers are more useful than one real screamer. None of my pocket calculators could handle such large numbers, so I tried the above formula on both sides of my main "computer" (using Quattro Pro/Dos, Prolog, Pascal, and the on screen calculator). In every case, the answer I got was 87,413.256955. I've also read reports on Compuserve from people who tested the formula on a 386DX+387 (using the 2.11 PMSPREAD spreadsheet, GWBasic and Rexx), an HP16C pocket calculator, a Sharp Business 10 Calculator, a child's Hello Kitty calculator, and a "handy-dandy calculator nerd watch". All returned the answer of 87,413.2569 (or some rounded variation). If one were to believe the LA Times, all of these calculating devices are defective!

Since I was unable to duplicate the LA Times' calculation on any device in my possession, I must assume that in preparing his article the author used the LA Times' "special news calculator" - the one they use to calculate all those questionable statistics and polls that fill the typical edition. However, as a public service to those who continue to believe that what's printed in the newspaper must be true, I am offering to accept free of charge any computers and calculating devices tested as defective according to the LA Times formula - thereby relieving the current owners of the terrible contradiction between their beliefs and reality.

Incidentally, for those running OS/2 on a flawed Pentium, there is a file available that will compensate for the hardware Pentium flaw in software and return the correct answer to the test formula. Check the OS2USER forum on Compuserve, or contact IBM.

Intel's cost for those 5 missing look-up table entries so far? End user lawsuits of more than $600 million, and sales of Pentium PC's dropped more than 20% in December. Kind of makes you wonder...