A Block Minifloat Representation for Training Deep Neural Networks