Power management has become a highly important focus in modern computing systems as energy is increasingly perceived as a critical resource. This has triggered the rise of topics related to energy-efficient computing. This paper presents an experimental study of three prevalent power management techniques: power limitation, frequency limitation, and ACPI/P-State governor modes (OS states related to power consumption). Using a benchmark with six computing kernels, we investigate power/performance trade-off with different hardware units and software frameworks (mainly TensorFlow and JAX). Our experimental results show that frequency limitation is the most effective technique for improving the Energy-Delay Product (EDP), which is a convolution of energy and running time. We experimentally observe that running at highest frequency compared to lowest one (resp. reducing frequency) might lead to an EDP at most 10× (resp. 1.68×) lower. Interestingly, while the frequency management behavior is transferable between different CPUs, we observe significant variations between TensorFlow (TF) and JAX, where the same power management settings sometimes produce opposite effects.