SWI-Prolog internally represents floats using the C-language type double. On most today systems this implies using a 64-bit IEEE representation. All floating point math functions are based on the C math-library.

The write/1 predicate and friends use the dtoa library by David M. Gay that prints floating point numbers with the minimal number of digits such that read/1 reads back the same value. If you want floats printed with a specific number of digits, use format/2:

?- A is pi, format('Pi = ~5f~n', [A]).
Pi = 3.14159
A = 3.141592653589793.

Floating point numbers are not exact. If you want exact arithmetic, please check out SWI-Prolog's support for rational numbers in the manual.

Quoting Richard O'Keefe:

SWI Prolog uses exactly the same floating point arithmetic as C. This is normally provided by your hardware, and can be expected to conform to the IEEE 754 standard. This is BINARY FLOATING-POINT arithmetic, not arithmetic on the mathematical real numbers. If you think you have found a mistake in Prolog's arithmetic, you are almost certainly wrong, and should make sure that you understand why, for example,

0.3*3 is not exactly equal to 0.9,

before you report the "bug".

See also

- What Every Computer Scientist Should Know About Floating-Point Arithmetic