Workload-Aware Power and Thermal Optimization for NVIDIA Jetson Processors using Advantage Weighted Regression

Abstract

Achieving power efficiency in embedded systems like NVIDIA Jetson processors requires balancing resource utilization, thermal management, and energy efficiency across workloads. A study proposes offline reinforcement learning (RL) for efficient system resource management without hardware interaction. By supporting CPU-intensive, GPU-intensive, memory-intensive, and hybrid workloads, energy efficiency and thermal stability improve. Advantage Weighted Regression, an offline reinforcement learning method, simplifies policy optimization. High-advantage actions are weighted more, guiding resource and power efficiency decisions. We optimize power adjustments and prevent overheating with a reward function to maximize system resource use. The method improves decision-making to mitigate noisy data-induced power changes in embedded systems due to inaccurate sensor readings and variable resource utilization. We compared the AWR-based policy to a static baseline method that cycles between increasing and decreasing load 50% of the time and randomly selects actions the rest of the time. In mixed workloads, the AWR-based policy reduces power consumption by 6.63%, CPU and GPU temperatures by 9.61% and 10.05%, swap memory usage by 20.67%, RAM utilization by 11.63%, and CPU utilization by 7.53%. CPU usage rises by 20.89% and RAM usage by 4.72%, respectively, during CPU-intensive tasks. Power consumption, GPU temperature, and swap memory usage in GPU-intensive tasks decrease by 17.14%, 17.35%, and 36.37%. Memory-intensive tasks reduce power consumption 17.6%, CPU temperature 6.07%, swap memory usage 52.06%, and RAM utilization 9.66%. Offline reinforcement learning with workload optimization outperforms non-adaptive power management at controlling temperature, using system resources, and saving power. This study recommends intelligent power management for energy-constrained embedded systems.