Phil Colangelo, Machine Learning Engineer, Intel Deep learning and artificial intelligence is quickly revolutionizing how we work and interact with technology in numerous ways. New algorithms and applications offer unprecedented accuracy, often exceeding human perception – however, as model sizes grow, the ability to adapt them for fast, power efficient performance limits the ability for these models to scale. This research explores the technical and practical feasibility of low precision deep neural networks, from 8-bit down to binary (1-bit) in for both activation and weights, and compares the accuracy of various low precision implementations across a standard neural network of various depths. These networks are then compared against theoretical and actual hardware implementations, using Intel Arria 10 and Stratix 10 FPGAs to create a hardware accelerator capable of greatly improved efficiency and throughput, with minimal accuracy loss relative to 32-bit floating point precision.