Rethinking floating point for deep learning

Jeff Johnson

Reducing hardware overhead of neural networks for faster or lower power inference and training is an active area of research. Uniform quantization using integer multiply-add has be…