Machine Learning for Energy Analysis

The Focus Group “Machine Learning for Energy Analysis” comprises as principal investigators Rudolf Diesel Industry Fellow Dr. Sani Nassif  (Radyalis)  and his host Prof. Dr.-Ing. Ulf Schlichtmann (TUM Department of Electrical and Computer Engineering).

Integrated Circuits are the key enablers all the “smart” devices we take for granted in our day-to-day lives. For example, the integrated circuit powering a modern smartphone has billions of components and represents the combined efforts of hundreds of designers and engineers. A key metric for such integrated circuits is the amount of energy they use, since that determines -for a smart phone for example- important factors such as battery life. Thus understanding the energy consumption of such circuits is crucial to creating efficient and marketable products. The prediction and management of energy consumption requires having an accurate model of the energy consumption as a function of environment (e.g. Temperature) and usage scenario. The complexity of these integrated circuits is such that it is often the case that some portions of the design are aquired from external design houses in order to save on development time. This presents a challenge when creating energy models for the design, however, since it is often the case that the details of such external portions are not known. This prsents a challenge for the prediction (during design time) and management (during use) of energy use.

This project aims to apply advanced machine learning techniques to aid designers solve this problem. We specifically leverage the fact that certain aspects of these energy models are known in order to develop "guided machine learning" techniques capable of creating accurate and predictive models useful in design and management. Note that this energy management is not specific to smart phones or battery operated devices. Global factors like climate change drive us to examine the energy use in all electronic devices that make use of integrated circuits. So these ideas and techniques can have the same impact on large data centers, which are known to use several percent of the total electricity generated.

TUM-IAS funded Doctoral Candidate:
Philipp Fengler, Electronic Design Automation