SANTA CLARA, Calif. -- Reports of a new floating-point bug involving Intel have hit the Internet. This time, the purported glitch involves Intel's high-end Pentium Pro and Pentium II processors.

"It's a bug in the floating-point integer store," said Robert Collins, who added that he plans to post details at noon EDT today on his "Intel Secrets" Web page.

"If you have a floating-point number and then try to store it as an integer, where the store would result in an overflow, the processor signals the wrong type of error," said Collins. "What the processor should do is signal an invalid-operand error; it does not."

According to Collins, the Pentium Pro and Pentium II chips will instead report that an overflow has occurred. Although this might seem to be the correct action, Collins said that the Pentium Pro manual notes that an invalid-operand error is in fact the proper response.

Intel said it was keeping a close watch on Collins's Web site. An Intel spokesman said the company called Collins and requested details on the purported problem, but Collins declined to provide any information in advance of his posting.

"We won't know what we've got until we see it on his Web site," said Intel spokesman Howard High. "We'll start investigating and characterizing it once we get the information.

High added that Intel may post a response at some point on the comp.sys.intel newsgroup. However, he said it might be some time before complete results of any investigation are available. He added that it's possible that the purported glitch may be one that's already been identified in the errata document for the Pentium II. (The Pentium II, which is a Pentium Pro that's been enhanced with MMX multimedia extensions, is due to be officially unveiled by Intel this Wednesday. The errata will be released at that point.)

This latest glitch is taking a somewhat tortured path to the Internet. Last week, Collins posted a message on the comp.sys.intel newsgroup that talked in general terms of a floating-point bug that has been reported to him by an unnamed college professor. A series of follow-up postings challenged Collins to back up his assertions, or retract them.

"Some people have been saying this is a hoax," said Collins. "It's at the point of no return--I have to post it." However, Collins still won't name the source of his information, other than to say it's an academic named "Dan."

Intel has already identified and made public several highly obscure anomalies involving floating-point math in the Pentium Pro. In the latest Intel documentation, errata number 2 is listed as "floating-point operand pointer may be incorrectly calculated after floating-point access which wraps the 64-Kbyte boundary in 16-bit code." Errata number 46 is "floating-point exception flag may not be set" in certain obscure scenarios. However, Collins said these items are not the flaw he's reporting.

The most similar existing errata, he said, is one that involves floating-point instruction store and is designated errata number 20 for the standard Pentium chip.

If the new floating-point reports spark message traffic on the Internet, it will conjure up memories of the Pentium FDIV bug of November, 1994. The flaw, which was first reported by EE Times, resulted in a public-relations disaster for Intel, its first-ever chip recall and an eventual charge against company earnings of $475-million.

Free Subscription to EE Times First Name Last Name Company Name Title Email address   Click here for your Free Subscription to EETimes Europe