Impact of Dynamic Voltage on GPU Energy Consumption for Real-Time Systems

Abstract Dynamic Voltage and Frequency Scaling (DVFS) has emerged as a crucial technique for balancing energy consumption and computational performance in CMOS-based systems. With the rise of highly parallel architectures like GPUs, applying DVFS effectively in real-time environments poses unique challenges and opportunities. By exploiting massive parallelism, high memory bandwidth, GPUs support real-time image and video in deep learning inference, meeting strict latency constraints required in real-time systems. This study investigates the application of DVFS on NVIDIA DGX systems equipped with Tesla V100 GPUs to balance energy consumption and deadline compliance for periodic task sets. By leveraging a necessary and sufficient schedulability condition, we dynam- ically adjust GPU frequencies to minimize energy usage while ensuring real-time constraints are met. Experimental results reveal a quadratic escalation in power consumption as task set sizes grow, necessitating higher operational frequencies to meet computational demands. Our findings underscore DVFS as a critical enabler for energy-efficient GPU computing in large-scale platforms, offering actionable insights for autonomous systems, and data centers where tasks present associated deadlines.