That isn't floating points fault, but whoever wrote the buggy parser.. Floating point is overall pretty good, sometimes I wish there was a version without some

That isn't floating points fault, but whoever wrote the buggy parser.. Floating point is overall pretty good, sometimes I wish there was a version without some many ways to store NAN(for 16 bit floats it is so wasteful), and maybe versions of sqrt/div/rcp/rsqrt that would clamp invalid inputs to 0, but other than that it is hard to see how you make any major improvements given the fixed # of bits. Actually one other possible improvement would be a way to request the hardware to dither when rounding down to the fixed bit count, although I'm not sure how it could do anything besides pure white noise, but better than nothing..