Subnormal

Subnormal floating point values (also known as denormal) are too small to be normalized yet can still be represented [1]. Depending on the specific computation, they may provide additional, useful precision. However, managing these values typically causes major slowdowns in a CPU's floating point unit [2].

For the Intel x86 architecture, floats between 1.17549429E-38 and 1.40129846E-45 are denormalized while for doubles the range is 2.2250738585072008D-308 and 4.94065645841246544D-324. [3]

For some CPU architectures, it is possible to control the treatment of subnormals. For instance, these values can automatically be rounded to zero.