@neauoire If you scale by 15/16 (which is very simple with integers!), you get an average better result:
https://flylib.com/books/en/2.729.1/high_speed_vector_magnitude_approximation.html