Paper Title
Machine Learning-Based Voltage and Frequency Control for Energy-Efficient GPU Computing

Abstract
Real-time, latency-sensitive cloud and edge performance requires energy-Efficient GPU execution. The critical problem is to maximize performance with minimal power consumption. The proposed predictive DVFS model in this paper dynamically scales frequencies of GPU before voltage changes. The Random Forest classifier, which qualifies CUDA runtime readings (kernel execution time, SM usage, and memory usage) is used to predict the optimal per-kernel frequency to lower the power consumption but not the performance. The framework is benchmarked against an Nsight Compute and gpowerSAMPLER on an NVIDIA Tesla T4 GPU on an AWS EC2 g4dn.xlarge instance with systematic testing over compute- and memory-intensive kernels. Performance-per-watt improvements are significant, with results up to 25 percent decrease in energy consumption, and run overhead of less than 3 percent. The software-based DVFS controller is proposed, which utilizes in-built GPU scaling capabilities, which allow flexible and scale-able energy-efficient processing in both cloud and edge architectures. Keywords - GPU energy efficiency, DVFS, Random Forest, CUDA profiling, performance-per-watt, AWS EC2, runtime inference, predictive modeling