🧮 October 1994. Number theory professor Thomas Nicely of Lynchburg College (Virginia) was studying the mutual values of twin primes. His Pentium was supposed to flawlessly execute millions of floating-point division operations. Instead, it spat out wrong results.
Nicely checked: 1 / 824633702441 — the result was wrong on three different Pentium machines. On older 486s — correct. He called Intel tech support. They told him: “Your math software license has probably expired.”
Yes. Intel tech support told a math professor his math license had expired.
⏱️ The bug lived in a processor component called the PLA (Programmable Logic Array) — a lookup table for division operations. Instead of calculating division “brute force” (expensive, slow), the Pentium used a precomputed table: the processor checked the table, got an approximation, refined it — and returned the result.
The problem: 16 entries in the table were missing due to a mathematical error in its definition. In 1994, Intel claimed a script had skipped 5 entries. But independent analysis later showed: 16 were missing, of which 5 actually triggered the bug, and 11 did not.
Intel had discovered this back in May 1994 during internal testing. Probability: one wrong result per 9 billion calculations. The company deemed it “a minor technical issue that doesn’t even qualify as errata” and quietly fixed the PLA in the next chip revision.
🔍 Intel knew. From May to October 1994, the company silently corrected the silicon in new processors, informing none of the buyers of the old chips. By October — when Nicely published his findings — millions of “defective” Pentiums were already in use worldwide.
Nicely sent an email to a dozen computer magazines. One recipient — Andrew Schulman, author of Undocumented DOS — forwarded it to Richard Smith. Smith posted the email on the Compuserve forum (a proto-social network of the 1990s). A reporter from Electronic Engineering Times spotted the post and, on November 7, wrote: “Intel fixes FPU glitch in Pentium.”
That could have been the end of it. But Intel decided to limit replacements — only customers who could convince an Intel engineer of their need for precision. If you couldn’t prove you required exact math — you were denied a replacement.
It was a strategic blunder the size of a small country’s GDP.
💥 IBM announced it was halting sales of all Pentium-based computers. It was a body blow — IBM was then the world’s largest OEM. The NASDAQ index crashed. Intel lost $10 billion in market capitalization in a single week.
Pentium users started calling the processor “Pentagone” — because it couldn’t divide accurately. Johnnie Carrey wrote a satirical FAQ: “How to check if you have the bug? Try dividing. If the result is wrong — you’ve got the bug. If the result is right — try again.”
Intel caved. The company announced a full replacement of any Pentium upon request — no proof needed, no persuasion required. Cost: $475 million. In its 1994 annual report, Intel admitted it had been “consumed by the controversy.”
🔧 The most interesting part is the cause. In 1994, Intel said: a script generating the lookup table had skipped 5 rows. But a die-photo analysis of the chip (conducted in 2024, 30 years later) showed: the real issue was a mathematical error in the table’s definition itself. The 16 rows weren’t skipped by the script — they were omitted by the mathematician who wrote it.
The difference is critical: a script can be rerun. Math needs to be rechecked. And no one at Intel rechecked it — because the PLA passed through a verification system that tested “a subset of all possible inputs,” not all of them.
The Pentium FDIV bug isn’t a story about math. It’s a story about how a tiny error in one table entry snowballs into a $475 million crisis through a chain of organizational decisions. Intel could have replaced all the chips in May — it would have cost peanuts. Instead, for three months, it hid the issue, restricted replacements, provoked IBM’s wrath, and lost $10 billion in market cap.
Every decision along the way — hide, limit, deny — was rational in the short term. And each one compounded the disaster in the long run. The bug was in the division table. But the real mistake was in Intel’s decision-making table.