Although a little indirect, the CUDA C Programming Guide does mention in the section on 16-bit floating point textures that the data type is “the same as the IEEE 754-2008 binary2 format”.
Also, according to the PTX manual, all half-float values (which they call f16 in the PTX manual) have to be upconverted to single or double precision floats before arithmetic can be performed on them, so the only relevant standards question is that of bit layout.