Energy Profiling of DNN Accelerators

This paper introduces a novel methodology for assessing the energy efficiency of neural network accelerators at both layer and network granularity. The approach involves extracting per-layer timing reports from recorded power profiles. The power and energy consumption of three prominent neural network accelerators, namely the Intel Neural Compute Stick 2, the Coral Edge TPU, and the NXP i.MX8M Plus is evaluated for three different Deep Neural Networks (DNNs) using this method. The study investigates the relationship between decreasing sampling frequencies and the average error, as well as the detailed energy consumption of individual DNN layers and layer types. The findings reveal that latency outperforms the number of operations per layer as a predictor for both overall and dynamic energy, with errors of 10% and 100% respectively. The main conclusions are: a sampling frequency of 200 kHz is necessary to achieve an average error of 5%; the number of operations is an inadequate predictor of energy consumption; and specific hardware settings significantly influence power and energy consumption, emphasizing the need for their consideration in estimation.