Float32 and Float16 output runs the risk of being cast inaccurately to integer datatypes by other software.