2 ms·
There are some interesting compatibility issues around 16 bit floating point types. As well as the IEEE layout (5 exponent, 10 mantissa) there’s bfloat16 which
by fanf2 6y ago
There are some interesting compatibility issues around 16 bit floating point types. As well as the IEEE layout (5 exponent, 10 mantissa) there’s bfloat16 which has the same range as 32 bit float (8 exponent) but less precision (7 mantissa) https://en.wikipedia.org/wiki/Bfloat16_floating-point_format https://en.wikipedia.org/wiki/Bfloat16_floating-point_format