Automated Training Data Construction using Measurements for High-Level Learning-Based FPGA Power Modeling

Abstract

Machine Learning (ML) is the process of developing Artificial Intelligence (AI) in computers, where the generated models are trained using appropriate learning algorithms and training data. For many machine learning techniques, especially the ones related to supervised methods, the construction of the training data highly affects the quality and accuracy of the derived model. In this paper w e present and evaluate an automated training set construction methodology where data is synchronously collected from both hardware and software. The complete design and data flow including the interaction between software and hardware, are thoroughly described. As a direct application, this work targets the construction of an FPGA-based circuit power modeling for subsequent early power estimation. The constructed Artificial Neural Network (ANN) model is trained using real measurement data sets extracted using a dedicated in-house designed and implemented generation and acquisition platform. The designated application falls under the power optimization area, becoming nowadays a major concern for most digital hardware designers, particularly in early design phases and especially in limited power budget systems. The power optimization approach in context can be extended in order to support online power management.