Navigating the Design Space of Reconfigurable Neural Networks Accelerators

M. Drumond,Parsa

Abstract:Neural Networks are an important class of algorithms used in many machine learning tasks, such as image classification and speech recognition. These algorithms are computeintensive and its users often need heterogeneous acceleration to achieve satisfactory performance. We survey the landscape of heterogeneous acceleration for Neural Networks, comparing three classes of accelerators, GPUs; ASICs; and FPGAs, according to three factors: flexibility, energy-efficiency, and scalability. We show that GPUs are flexible but suffer from poor energyefficiency, ASICs are energy efficient but inflexible, and FPGAs can achieve the flexibility of GPUs and near ASIC energyefficiency.