Efficient operation of high-performance computing (HPC) facilities requires careful control of energy consumption, which largely depends on the computational workload and the types of code being processed. Modern processors prioritize higher core counts and clock speeds to maximize performance, but this approach leads to increased energy usage in proportion to the processing demand. In this article, we explore the use of undervolting—a method of limiting the processor power and clock speed—to balance computational performance with energy efficiency. Using machine learning-based predictive modelling, we analyse historical power consumption data from the University of Guadalajara’s HPC system, Leo Atrox, to model energy usage patterns across different tasks. Our findings demonstrate that task-specific power consumption is nonlinear, and that strategically restricting power based on workload requirements can significantly optimize the overall energy efficiency of an HPC system.