Resource-demand Estimation for Edge Tensor Processing Units

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