damerell@nsfnet-relay.ac.uk (RHBNC) (05/17/89)
There has been a lot of discussion about the speed of floating-point calculations, but I havent seen any mention of accuracy. Suppose you have two routines, both called "sin": 1: sin(x) as accurately as the hardware allows, 2: sin(x) with error about 100 x the error of 1. I would expect 2 to run about twice as fast as 1, besides being easier to program. But CPU time is cheap and wrong answers can be very expensive. I maintain that any worth-while bench test must check the accuracy of the results as well as speed. mark.