The Hidden Power of Pure 16-bit Floating-Point Neural Networks