An in-depth study on 8-bit floating-point number formats for activations, weights, and gradients - 42Papers