Would you trust your life to a Pentium? How about a neural network?
Raphael Needleman
I suppose, as the editor in chief of this magazine, somewhere in my job description there's a line that says, ``...shall pontificate about every major screwup committed by Intel, IBM, or Microsoft.'' Of course, if I did that, I'd never have any time left to write about the things that computer companies get right, but nonetheless, I feel strangely compelled to say a few words about the dearly departed floating-point bug in the Intel Pentium.
So here it is, my advice to everybody who's been trying to figure out how serious the Pentium bug is, how Intel messed up, and whether we should be buying other chips like PowerPCs (or 486s) instead: Lighten up! Buy a Pentium if you want. At least now you know about one of its bugs. Really, what do you expect? T
he Pentium chip may look and feel like hardware, etched as it is in silicon, but it's just software that happens to have been pressed into hardware. And name me one complex program that doesn't have a bug.
Now, granted, we all learned a lot about the computer industry because of this bug. Mostly, we learned that no matter how minor or technical the error, once the popular media get hold of it, the company responsible is in serious public-relations trouble. Remember the ridiculous ``unintended acceleration'' fiasco that nearly sunk Audi in the U.S. a few years ago. For me, the exact time that the Pentium bug became part of the popular culture was one cold December morning. At 8:15 a.m., I heard a Chevys Mexican restaurant advertisement on the radio lampooning the chip's bug (Chevys' tortilla chips, it is claimed, are not in need of a recall).
The popular media may not be aware of it, but there are doubtless other undiscovered bugs lurking in the Pentium, not to mention its competitors. Can we eve
r expect these flaws to be eradicated?
Not really. Simulations are one thing, but we all know that there's no real way to run every possible combination of inputs into a CPU to find those that work right and those that don't.
From the pedestrian Pentium bug, I want to make a giant leap to another idea, one that has only a little to do with the Pentium but everything to do with the way new technology is designed today. Here goes: The nature of machine design means that each new machine can carry with it the inherent flaws of the generations of machines that came before it and contributed to its design.
It's pretty easy to tell when a simple machine built by a simple tool has a flaw (e.g., an off-kilter table built using a flawed level), but as machines get more complex, how do we make sure the chain of tools and designs runs true? Is it possible that we could be living with evil recessive lines of compiler code that will reach out and bite us sometime in the future, several generations of
code removed from the original bug?
Fortunately, as we build new machines using older tools, we're also creating debugging equipment that can catch dormant problems because it's smarter than the last-generation designs--we hope. And even though I don't think the problem of machine evolution is serious at this moment, as we use more heuristic methods of programming and more neural networks and as we become satisfied with programs that do what we want even when we don't know why, we are just asking for trouble. Let me put it this way: Although you might be happy to entrust part of your stock portfolio to a neural-network algorithm that's outperformed the Dow, how would you feel if the airliner you were on was programmed using heuristic methods?
The Signal in the Noise. Speaking of trouble, I'm embarrassed to report that a number of readers noticed my improper use of the phrase ``high signal-to-noise ratio'' in last month's column. What I was trying to allude to was the high junk factor in broadcas
t data, which is, of course, an illustration of a low signal-to-noise ratio. I stand corrected.
Finally, if you're interested in a thorough explanation of the exact nature of the Pentium bug, turn to ``The Truth Behind the Pentium Bug'' for Tom Halfhill's excellent and frightening account of how the flaw came to be--and why it was so easy for Intel to correct it.
Raphael Needleman, Editor In Chief, (
rafe@mcimail.com
), fax (603) 924-2550
Mutant Chips
