In RISKS, Kahan was claimed to have said that INMOS got the spec
wrong -- I have no idea what he meant or whether he would agree with
this report.
>5) The Pentium bug seems to have been caused by "missing entries"
>in a division lookup table. I'm fairly certain that our production
>testing used evaluations that would exercise every entry in our
>division and multiply lookup tables - we caught some yield hazards
>in revB that way (multiply table outputs were not fast enough in
>75% of chips!) - if Intel had used this type of testthen surely
So how much of the verification was testing and how much was mathematical
proof?