Understanding the Pentium Disaster

When Tarun Soni, a contract space agency software engineer, read on Internet about a floating point flaw on Intel Corporation's Pentium microprocessor, he checked his Pentium-powered home computer, a new Gateway.

"The moment I said I work for NASA, they said, 'If you do NASA work, you need the chip replaced,'" he said. "I found their damage control much better than I expected" after reading Internet and press reports.

Damage control indeed.

Soni is just one of thousands of personal computer buyers getting a replacement chip as the leading PC semiconductor supplier wrestles with a public relations debacle over the processor's failure to do certain floating point divisions beyond four decimal places. Intel engineers discovered the flaw during ongoing random testing last summer, 14 months after the chip went to market.

Since the flaw evaded the most exhaustive testing the semiconductor supplier had done on a product, they concluded an average user would get a failure once in 27,000 years. They considered the bug so obscure, they did not note it on their errata sheet to users, notify their technical support group or halt distribution.

The debacle worsened last week when IBM Corp. announced a halt in the middle of the holiday buying season to sales of Pentium-powered computers until a reliable chip can be installed.

The firm concluded that an average spreadsheet user could expect a glitch every 24 days. IBM also pledged to replace the flawed unit on demand, in stark contrast to Intel's we'll-replace-if-we-think-you-qualify policy. IBM makes a competing microprocessor, the PowerPC, and only 5 percent of its PC sales are Pentiums.

Still, whatever IBM's motives, it did succeed in forcing Intel to relent. On Dec. 20, Intel finally announced it would replace every flawed Pentium chip, a move that could cost the firm hundreds of millions of dollars. A big hit, to be sure, but one the semiconductor giant can absorb.

But that still leaves larger questions. How could Intel have botched its PR effort so badly, if the problem has been grossly exaggerated? How common are such glitches -- in any microprocessor, stuffed as they are with millions of transistors and complex designs?

There is no clear, definitive answer to these questions -- one reason the controversy, perhaps, has people so worried. Oracle Government of Bethesda, Md., for example, reviewed the problem after it received customer queries. "Calculations in our database [program] are not affected," said Ken Stewart, a senior systems engineer. "We use the fdiv [long division, which was thought to product the error] instruction only to get an initial estimate for calculations, then use our own algorithms. We go 38 decimal places, and the floating point of any microprocessor doesn't give us the precision we need.

"Once you get it out of Oracle, though, if you do numerical analysis with the Pentium, there may be a problem," he warned.

Others put a plague on both IBM's and Intel's houses. "Intel's estimate of 27,000 years and IBM's estimate of 24 days both are not worth the paper they're written on; and they were both done by friends of mine," says William Kahan, an electrical engineering and computer science professor at University of California at Berkeley who is renowned for devising tests for just such problems. Kahan feels IBM's stand is "an overreaction," but advises Pentium users "not to rely on anyone else's estimate of the risk," including the damage that might be done replacing the microprocessor. "Some people should change their chip," he says, "but many people use software that doesn't even have floating point in it."

Nicely feels IBM "threw more gasoline on the fire. I find their assessment... far-fetched. If IBM's interpretation is correct, thousands should have reported the flaw."n

But he also blames Intel for creating its own monster by departing from established procedure.

Still, some damage may be irreversible. Even as Intel amasses the support of more than a dozen computer manufacturers to blunt IBM's blow, the firm faces a dozen lawsuits and warnings by a technology consultant to Fortune 500 companies, and a trade organization, against buying the machines.