number of exponent bits. bfloat16 has a larger dynamic range with lower precision. e.g. this would be infinity in fp16 https://float.exposed/b0x5f80