The IEEE Standard 754 specifies values and requirements for floating-point representation (such as base 2). The standard outlines requirements for two formats: basic and extended, and for two word-lengths within each format: single and double.
Intel Fortran supports single-precision format (REAL(4)) and double-precision format (REAL(8)) floating-point numbers. At some levels of optimization, some floating-point numbers are stored in register file format (which equals 1 bit sign, 15 bit exponent, 64 bits of significand rounded to 53 bits), rather than being stored in IEEE single or double precision format. This can affect the values produced by some computations. The compiler option -fp-model (Linux* and Mac OS*) or /fp (Windows*) can control how floating-point expressions are evaluated, thus leading to more predictable results.