Search for a command to run...
Fine-grained power estimation is fundamental to modern MPSoC design and operation, enabling early power and thermal budgeting before fabrication, predictive thermal control at runtime, post-silicon microarchitectural energy debugging, and hardware security analysis. Existing System-on-Chip (SoC) designs often rely solely on total power consumption measurements because of the cost, size, and complexity of power sensors. However, the lack of precision and granularity in power measurements leads to inefficient power and thermal management and limits the ability to detect failed or compromised sensors. To tackle these challenges, we propose the Improved Clustering Blind Power Identification (ICBPI) approach, an innovative approach for fine-grained power estimation in SoCs. ICBPI leverages total power measurements and per-unit temperature data, even with potentially unreliable sensors, to deliver accurate per-unit power estimates. This dual focus not only enhances power estimation, leading to better power management decisions, but also improves the detection of malicious sensors—such as those compromised by hardware Trojan attacks—that manipulate sensor data to induce potential system failures. Extensive evaluations across four SoC architectures demonstrate ICBPI’s superior performance in reducing per-core power estimation errors by \({77.56\%} \) compared to the baseline Blind Power Identification (BPI) and by \({68.44\%} \) compared to the state-of-the-art approaches. Furthermore, ICBPI achieves \({100\%} \) detection of malicious thermal sensor attacks and enhances the localization process compared to the Blind Identification Countermeasure (BIC) approach. In the comparative analysis between ICBPI and the most recent model, Alternating Blind Power Identification (ABPI), ICBPI shows error rates as low as \({0.5\%} \) compared to \({2.8\%} \) in a 2 × 2 mesh benchmark and outperforms ABPI on heterogeneous architectures. The NVIDIA Jetson Xavier AGX development board serves as a platform for validating ICBPI. Consequently, ICBPI demonstrates an approximate reduction in average error of \(79.71\% \) on the CPU and \(79.59\% \) on the GPU relative to three existing state-of-the-art approaches.