ML-based Power Estimation of Convolutional Neural Networks on GPGPUs

The increasing application of Machine Learning (ML) techniques on the Internet of Things (IoTs) has led to the leverage of ML accelerators like General Purpose Computing on Graphics Processing Units (GPGPUs) in such devices. However, selecting the most appropriate accelerator for IoT devices is very challenging as they commonly have tight constraints e.g., low power consumption, latency, and cost of the final product. Hence, the design of such application-specific IoT devices becomes a time-consuming and effort-hungry process, that poses the need for accurate and effective automated assisting methods.In this paper, we present a novel approach to estimate the power consumption of CUDA-based Convolutional Neural Networks (CNNs) on GPGPUs in the early design phases. The proposed approach takes advantage of a hybrid technique where static analysis is used for features extraction and the K-Nearest Neighbor (K-NN) regression analysis is utilized for power estimation model generation. Using K-NN analysis, the power estimation model can even be created with small training datasets. Experimental results demonstrate that the proposed approach is able to predict CNNs power consumption up to a Absolute Percentage Error of 0.0003% in comparison to the real hardware.