Accurate Load Prediction in Dynamic Voltage Frequency Scaling Systems

Dynamic Voltage and Frequency Scaling (DVFS) is a potent technique employed to minimize dynamic power consumption in Silicon on Chips (SoCs). The accurate determination of future loads is pivotal in DVFS systems, as it influences the selection of desired frequencies and optimal voltages while ensuring circuit performance is not compromised. In this paper, we introduce a novel approach and algorithm for precise future load determination in processors for DVFS systems using Machine Learning Time Series Method. Our proposed algorithm dynamically evaluates various models from the model repository, considering different input parameters to identify the optimal model for predicting future loads. The proposed approach is validated using simulation taking real server data of AMD Ryzen™ 9 7900X3D processor. This proposed approach and algorithm yields significantly more accurate load predictions with selection of the Neural Network Model, out of four popular models, results in the fastest convergence time and achieves a minimal Mean Absolute Percentage Error (MAPE) of 0.133. This proposition effectively reduces the power consumption enhancing the efficiency and effectiveness of DVFS systems by tailoring load predictions to processor workloads, ultimately leading to improved power management in SoCs.