This paper examines the performance of two power efficient hardware implementations using deep neural networks to perform a simple image classification task. We provide the first ever examination of the accuracy-energy trade-offs of deep neural networks running on both an embedded GPU, and a neuromorphic processor. IBM's TrueNorth is a brain-inspired event-driven neuromorphic processor. It was designed to be scalable and to consume extremely low amounts of power. NVIDIA's Tegra K1 SoC is a mobile processor also designed with low power and a small footprint in mind. While these two chips were designed with similar constraints, the resulting architectures and performance trade-offs achieved are significantly different. On our simple image classification task Convolutional Neural Networks utilizing the Tegra K1 SoC achieve up to 89 % accuracy with a normalized accuracy per active energy, ||Acc||/EA, score of up to 24.22 on our test dataset, while Tea Networks running on the TrueNorth processor achieve less accuracy at 82%, but a better accuracy-energy trade-off with a ||Acc||/EA score of up to 158.49.

Abstract not provided