High Performance Neural Networks with Small Floating Point (WP530) - This white paper introduces a 7-bit small floating point (SFP) number representation that can implement deep neural network models with the same accuracy as INT8 but with 60% higher performance for ResNet-50.