Power and Performance Analysis of Deep Neural Networks for Energy-aware Heterogeneous Systems

The driver of technology innovation is shifting from raw computing performance to performance delivered per watt. Therefore, it is crucial to conduct heterogeneous (CPU-GPU) system performance analysis in terms of power utilization. The main objective of our experimental study is to provide a detailed analysis of performance and power utilization of Convolution Neural Network for image classification of CIFAR-10 tiny images. We present an approach to calculate one convolution-layer power utilization for heterogeneous CPU-GPU systems by employing CUDA and OpenCL environments. The purpose of power, performance and hardware utilization analysis is to promote green computing and to assist system designers and AI specialists in choosing a green neural network architecture for energy-aware high-performance heterogeneous systems.