Floating-point architectures (continued)
Divide and square root sometimes replaced by reciprocal approximation and Newton-Raphson iteration (some models of Cray, and HP/Intel IA-64), and result may be incorrectly rounded
Intel x86 and IA-64 do not produce correct rounding (1 bit error on Pentium, 0.6 bit error on IA-64: helped by 80/82-bit format)