Mathematician Finds Intel’s Pentium Doesn’t Compute : Technology: A flaw that the company failed to disclose in June causes errors in complex calculations.

A mathematician in Lynchburg, Va., was the first to get bugged by a flaw in Intel Corp.’s popular Pentium chip.

While doing complex division problems, Thomas R. Nicely, a mathematics professor at Lynchburg College, detected a defect that caused his department’s three Pentium-based computers to spew out faulty numbers in the ninth place to the right of the decimal point and beyond. So he called Intel a month ago, and now he’s garnering a Warholesque modicum of fame as the man who blew the whistle on the chink in Intel’s famous chip.

Interestingly, Intel’s own engineers, doing random calculations, had themselves discovered the imperfection in June. Yet the company chose not to disclose the problem to customers, a spokesman said Wednesday, because it viewed as minuscule the chance that anyone other than an advanced research scientist would ever encounter it.

“The average user is not going to see this thing, except maybe once in 27,000 years,” said John Thompson, a spokesman for the Santa Clara-based chip maker, the world’s largest.

Thompson added that the bug has been fixed as part of a “scheduled improvement” in new Pentium microprocessors that began shipping two months ago. (To fix the problem, Intel had to add a small number of transistors to the thumbnail-size Pentium, which contains 3.3 million transistors). The company is also offering to replace troublesome chips should customers call and demand new ones. For purchasers buying in quantity, those high-powered chips run anywhere from $383 to $935 each.

Intel stock closed up 37 1/2 cents at $65.125 a share Wednesday.

Some in the industry were not so quick to sweep the silicon dust under the rug. Microprocessors, after all, are commonly known as the “brains” of personal computers--in other words, the component that makes them tick.

“The consequences are serious because you can’t trust it,” said Cleve Moler, the co-founder and chief scientist of MathWorks Inc., a software company that was touting its solution to the problem Wednesday. “Nobody’s gotten a wrong answer yet, but everybody’s worried.”

MathWorks, based in Natick, Mass., said has begun distributing to its customers a modification of its $1,700 MATLAB technical computing program that works around the bug. Moler also plans to post a description of the technique--which basically entails having the computer check its own work--on the Internet, where news groups have been churning with complaints and concerns about the bug for weeks.

*

The problem was staying nicely under wraps until Prof. Nicely called Intel’s tech support in late October. As usual in such matters, he said, the folks answering the phone had no idea what he was talking about. But eventually the problem filtered up to Intel engineers, Nicely said, and “it got everybody’s attention.”

Nicely had first noticed problems with his calculations in June after he had been running the math department’s three Pentium systems night and day for several months trying to check certain properties of prime numbers and all the “prime constellations”--double, triple and quadruple prime numbers, for example. (A prime number is an integer that can be evenly divided by no other whole number other than 1 and itself, such as 2, 3, 5 or 7.)

It was in analyzing a property of twin prime numbers (such as 1113) that the bug manifested itself, Nicely said.

Nicely, 51, who has been teaching math at his 101-year-old liberal arts college since 1968, was out to break the world record in this particular exercise in pure math, which previously had been done on supercomputers. The idea was to demonstrate that powerful new desktops could do the work even better.

At first, he said, he attributed a discrepancy with past calculations to an error in his software program. Then, he figured it had to be the program that compiled the numbers. But after four months he finally fingered the least likely suspect: the chip itself.

Having to stop his calculations at the number 6 trillion and sort out the trouble “was a nuisance for me,” he noted, with perhaps surprising understatement.

Since then, Intel has furnished him with three new Pentium chips. After 1 quadrillion calculations, so far all appears to be correct. Nicely, who describes himself as “one of the obscure mathematicians in the herd,” said Intel has invited him to be a consultant on this type of problem; he also hopes to get a few machines donated to the college.

*

This is not the first such trouble for Intel. In the early 1980s, it recalled the first runs of its 386 chip, but only a few thousand were in distribution at the time. Millions of Pentiums are on the market, however.

For now, Nicely is pleased that this particular problem seems to be solved, although, he wondered, “who knows what other bugs are lurking?”

In the meantime, he is warning his students not to treat computers “as articles of faith.”

Separately, it was reported that both the Justice Department and the Federal Trade Commission plan to investigate anti-competitive practices at Intel.