Shifted and Squeezed 8-bit Floating Point format for Low-Precision Training of Deep Neural Networks

Cambier, Léopold, Bhiwandiwalla, Anahita, Gong, Ting · arXiv · 2020