Artificial Intelligence (AI) has changed our daily lives. The evolution from centralised cloud-hosted services towards embedded and mobile devices has shifted the focus from quality-related aspects towards the resource demand of machine learning. Its pervasiveness demands for "green" AI-both the development and the operation of AI models still include significant resource investments in terms of processing time and power demand. In order to prevent such AI Waste, this paper presents Precious, an approach, as well as practical implementation, that estimates execution time and power draw of neural networks (NNs) that execute on a commercially-available off-the-shelf accelerator hardware (i.e., Google Coral Edge TPU). The evaluation of our implementations shows that Precious accurately estimates time and power demand. CCS CONCEPTS• Hardware → Power and energy; • Computer systems organization → Embedded systems.

Machine learning has shown tremendous success in a large variety of applications. The evolution of machine-learning applications from cloud-based systems to mobile and embedded devices has shifted the focus from only quality-related aspects towards the resource demand of machine learning. For embedded systems, dedicated accelerator hardware promises the energy-efficient execution of neural network inferences. Their precise resource demand in terms of execution time and power demand, however, is undocumented. Developers, therefore, face the challenge to fine-tune their neural networks such that their resource demand matches the available budgets. This paper presents Precious , a comprehensive approach to estimate the resource demand of an embedded neural network accelerator. We generate randomised neural networks, analyse them statically, execute them on an embedded accelerator while measuring their actual power draw and execution time, and train estimators that map the statically analysed neural network properties to the measured resource demand. In addition, this paper provides an in-depth analysis of the neural networks’ resource demands and the responsible network properties. We demonstrate that the estimation error of Precious can be below 1.5 % for both power draw and execution time. Furthermore, we discuss what estimator accuracy is practically achievable and how much effort is required to achieve sufficient accuracy.

No abstract