BFloat16[1] is an interesting tweak to the original, but built with more modern requirements for ML in mind. Particularly because it is easier to think of 32 b

BFloat16[1] is an interesting tweak to the original, but built with more modern requirements for ML in mind. Particularly because it is easier to think of 32 bit IEEE-754 to BFloat16, but with fewer bits of think about (& possibly you can enumerate the entire range in a laptop to test something like "will this function work for all values?"). [1] - https://en.wikipedia.org/wiki/Bfloat16_floating-point_format#Contrast_with_bfloat16_and_single_precision https://en.wikipedia.org/wiki/Bfloat16_floating-point_format...