In 1994, users of Intel’s then-new Pentium processor discovered that the chip could return slightly wrong answers for certain floating-point divisions. The error was rare and data-dependent, but it was real: for a small set of operand combinations the divide instruction produced a result with reduced precision rather than the correct value. The flaw became known as the FDIV bug, after the x86 floating-point divide instruction.
The cause was a defect in a lookup table. The Pentium’s floating-point unit divided numbers using a radix-4 SRT algorithm (named for Sweeney, Robertson, and Tocher), which generates two quotient bits per clock cycle by consulting a table of precomputed values stored in a programmable logic array. The table should have held 1066 entries, but because of an error in the script that loaded the array, five entries were left out. When the divide hardware indexed into one of those missing cells, it fetched zero instead of the correct value, and the resulting quotient was wrong.
Intel’s own response is documented in its in-house study, “Statistical Analysis of Floating Point Flaw in the Pentium Processor,” by H. P. Sharangpani and M. L. Barton of Intel Corporation, released in late 1994. The paper argued that the flaw was extremely unlikely to affect ordinary users, estimating that a typical user would encounter an error only once in many thousands of years of normal computing. Intel initially treated the bug as a minor erratum and offered replacements only to customers who could demonstrate they were affected.
That position collapsed under scrutiny. The flaw had been found independently by mathematician Thomas Nicely while computing sums over prime numbers, and once the story spread, others pressed harder on Intel’s statistics. As Alan Edelman’s analysis “The Mathematics of the Pentium Division Bug” explains, the relevant question was not how often a random division fails but how often real software, which does not divide by random numbers, would hit one of the bad table entries. IBM researchers argued that some accounting workloads could encounter the error far more often than Intel’s once-in-millennia figure suggested.
Facing mounting public pressure, Intel reversed course in December 1994 and agreed to replace any affected Pentium on request, no questions asked. The company took a charge of roughly 475 million dollars to cover the recall and replacement of the flawed chips. The fix itself was simple: later Pentium steppings loaded the complete table. The lasting lesson was about handling: a narrow, well-understood hardware defect became one of the most famous failures in computing history less because of the silicon than because of how its maker first chose to talk about it.