Floating Point Numbers

• IEEE 754 is the most popular one
• bfloat16 offers same range as IEEE 754 32-bit floats (single float) but reduces accuracy in the significand. It is used in ML, and available in M2 chips, Intel chips, TPUs, NVIDIA GPUs, etc.
• Common Lisp doesn't specify float format for implementations but only specifies minimum accuracy (in bits) for exponent and precision [HyperSpec]
• See also: Unum