High-Precision Hybrid Power Modeling for Virtual Prototype Platforms Based on Machine Learning

Energy efficiency is essential in system-on-chip (SoC) design, with early-stage power analysis crucial for effectively exploring the design space. However, current methods struggle to accurately predict power consumption for multi-modules in the SoC in high-level simulations. This paper introduces a hybird power modeling framework based on machine learning methods, which can be integrated into the RISC-V virtual prototype (VP) to provide accurate power estimation at the electronic system level (ESL). First, this methodology supports the modeling of both leakage and dynamic power for the three major modules of a SoC (CPU, bus and memory). Subsequently, a broad set of modeling features is identified based on the working mechanism of the RISC-V VP, followed by the application of sequential forward selection. Finally, advanced regression techniques are applied to enhance the accuracy of the modeling results. The dynamic power models for the CPU, bus, and memory modules achieved mean absolute percentage errors (MAPEs) of 5.81%, 6.75%, and 16.90%, respectively, while the leakage power models exhibit MAPEs consistently below 5%.