Comment by JoeAltmaier
3 years ago
I remember being surprised when Intel released a Pentium chip that couldn't do floating point right. It would have been trivial to, I don't know, try dividing by 10 and then multiplying back and checking the answer. It couldn't do that.
At the time I was thinking, their test group could have tested every single mantissa over several operations exhaustively. Why didn't they? What excuse could they have had? It was a computer; they feel no pain when asked to do laborious repetitive tasks.
You'd think so, except no. Floating point operations are innately lossy and even non-associative, and non communative.
A+B+C does not equal B+C+A in floating point.
Which should happen in entirely predictable and orderly fashion. In fact, they should be testing for that too.