Unlocking Efficient Power Modeling with Machine Learning

In this article, we propose a novel approach to modeling power consumption in out-of-order CPU cores, which combines the strengths of traditional analytical methods and machine learning (ML) techniques. The proposed method, called PANDA, aims to overcome the limitations of existing approaches by unifying both analytical and ML solutions.
Traditional analytical methods are accurate but oversimplify the complex relationships between hardware components, while ML models can capture these relationships but struggle with unknown new configurations. PANDA addresses these issues by integrating the best of both worlds, leveraging the background knowledge of CPU architecture design to create a unified model that can adapt to new configurations.
PANDA’s resource function F i res of each major component are derived based on the architecture-level power modeling methods and are tailored to the target out-of-order CPU core. These functions capture the unique characteristics of each component, such as fetch width, decode width, and cache size, which significantly impact power consumption.
PANDA’s hybrid approach utilizes both analytical and ML techniques to model power consumption. The analytical methods provide a solid foundation for the model, while the ML techniques enable adaptability to new configurations. This unified approach enables accurate predictions of vector-based power values for each specific workload.
In summary, PANDA offers a novel approach to power modeling in out-of-order CPU cores that combines the strengths of traditional analytical methods and ML techniques. By leveraging background knowledge of CPU architecture design and utilizing both analytical and ML techniques, PANDA can adapt to new configurations and provide accurate predictions of vector-based power values for each specific workload. This unified approach has the potential to significantly improve the accuracy and efficiency of power modeling in out-of-order CPU cores.