3 ms·

I mostly agree. But bfloat16 at least uses 16 bits. tfloat32 is a 19 bit format and that just seems like a terrible naming decision to me.