In October 1994, Dr. Thomas Nicely, a mathematics professor in Virginia, noticed something bizarre while calculating twin primes on a brand new Intel Pentium computer: division was giving the wrong answer. A flaw in silicon meant \(4,195,835 / 3,145,727\) produced \(1.333739\) instead of \(1.333820\). That single hardware error ended up costing Intel $475,000,000 — all because of 5 missing numbers in a silicon lookup table.
Dr. Thomas Nicely was computing Brun's constant — the sum of the reciprocals of all twin primes:
To verify his calculations, he ran the code across multiple computers: 486DX, 386, and a brand-new 66 MHz Pentium machine. The 486 and 386 agreed to 19 decimal places. The Pentium diverged at the fifth decimal place.
Dr. Nicely spent six months ruling out software bugs, memory errors, compiler flags, and motherboard defects. Finally, he isolated the simplest possible reproducible expression:
Why did Intel design the division unit this way? To understand the bug, you must understand how silicon calculates division.
In standard long division, hardware guesses 1 quotient bit per cycle: compare divisor to remainder, subtract if it fits, shift. For a 64-bit double, this requires 60+ clock cycles — agonizingly slow for 3D graphics and engineering simulations.
The Pentium introduced a high-speed SRT Radix-4 division engine (named after Sweeney, Robertson, and Tocher). By computing in Base-4, it generated 2 quotient bits per cycle, cutting division time down to just ~19 cycles!
In Radix-4, the quotient digit \(q_i\) is chosen from a redundant digit set:
To implement the quotient selection logic in silicon, Intel mapped the mathematical Robertson diagram into a Programmable Logic Array (PLA). Out of the 2,048 possible input combinations, only 1,066 combinations were mathematically reachable.
A software engineer wrote a simple C script to download the mathematical table entries into the silicon mask layout file. The script contained an off-by-one loop error:
// A C loop copied table entries into the silicon PLA bitmask: // Due to an off-by-one array index boundary, the loop halted early: for (int row = 0; row < 1061; row++) { // ❌ SHOULD HAVE BEEN 1066! write_pla_row(row, srt_table[row]); } // Exactly FIVE entries at the boundary were never written!
Because the C script stopped at row 1,061 instead of 1,066, exactly 5 entries were never written into the silicon mask.
In physical CMOS transistor logic, an unprogrammed bit cell in a PLA defaults to logic 0 (Ground). So when a calculation hit one of the 5 missing entries, the CPU didn't halt or throw an exception. It quietly selected quotient digit \(q_i = 0\) instead of \(+2\).
Because it selected \(0\), the hardware failed to subtract \(2 \times \text{divisor}\) from the remainder. The partial remainder grew out of bounds, poisoning all remaining 50+ bits of the division calculation!
When Dr. Nicely posted his findings to the comp.sys.intel Usenet forum on October 30, 1994, Intel initially tried to downplay the issue:
"An error only occurs once in 9 billion random floating-point divisions. The average spreadsheet user would only encounter it once every 27,000 years."
The engineering community was furious. A researcher named Tim Coe reverse-engineered the exact PLA table and proved that in engineering, CAD, financial modeling, and scientific simulations, certain iterative algorithms hit the vulnerable entries repeatedly.
On December 12, 1994, IBM dropped a bombshell: it halted all shipments of Pentium-based personal computers worldwide. Intel's stock cratered.
On December 20, 1994, Intel CEO Andy Grove surrendered and announced an unconditional, no-questions-asked replacement policy for all Pentium chips. Intel took a $475 million pre-tax charge against its earnings — roughly half of its entire profit for the fourth quarter of 1994 (equivalent to over $1 Billion today).
Unlike software, where a bug can be patched with an overnight over-the-air update, silicon is physical. Once a billion transistors are etched in glass and silicon, there is no "Ctrl+Z".
The FDIV disaster permanently altered the VLSI engineering landscape in three massive ways:
// In modern ASIC design, table completeness is formally proven: property p_srt_lookup_complete; @(posedge clk) disable iff (!rst_n) (fdiv_state == STATE_QUOTIENT_SELECT) |-> !$isunknown(pla_table_out[quotient_index]); endproperty assert_srt_no_unwritten_cells: assert property(p_srt_lookup_complete) else $fatal(1, "FORMAL FAULT: Unmapped cell accessed in SRT division PLA!");
Modern CPUs are verified using Formal Verification Tools (like Cadence JasperGold and Synopsys VC Formal). Instead of testing random division vectors, formal tools use mathematical solvers to prove that the RTL circuit matches an ideal mathematical specification for every single one of the \(2^{128}\) possible input combinations.
Intel recycled thousands of returned processors into keychains and novelty desk ornaments with the inscription: "Bad chips made into good keychains." Some were also repurposed into non-floating-point embedded products where the FPU was physically disabled.
The bug triggers when the partial remainder \(r\) and divisor \(d\) fall into specific ranges (specifically where the scaled divisor is close to \(1.444\) and the remainder lands in the top boundary of row 1,061..1,065). In binary, these are values where the divisor has bit patterns with many repeating ones (e.g. \(3,145,727 = 0010\ 1111\ 1111\ 1111\ 1111\ 1111_2\)).