Benchmark-based Study of CPU/GPU Power-Related Features through JAX and TensorFlow

TL;DR

This study addresses energy-aware computing by evaluating three power-management techniques—DVFS (frequency limitation), power capping, and ACPI/P-State governors—across CPU/GPU platforms using a six-kernel benchmark and two Python frameworks, TensorFlow and JAX. It employs the EA2P profiler to measure time, energy, and the Energy-Delay Product $EDP = E imes T$, revealing that frequency limitation often yields the best $EDP$ improvements, though gains are highly workload- and platform-dependent. Results show distinct, framework-specific behaviors: on Nvidia A100 GPUs, power capping delivers substantial $EDP$ reductions with modest timing penalties; on Intel, power capping can be highly effective while DVFS can incur notable time costs; and on AMD, DVFS frequently provides the lowest $EDP$ for many workloads, albeit with trade-offs. The work emphasizes the need for platform- and workload-aware energy management policies and highlights differences in framework robustness (e.g., JAX stability at very low DVFS on A100) and memory-management strategies, offering guidance for practitioners, framework developers, and hardware designers to optimize energy efficiency in high-performance settings.

Abstract

Power management has become a crucial focus in the modern computing landscape, considering that {\em energy} is increasingly recognized as a critical resource. This increased the importance of all topics related to {\em energy-aware computing}. This paper presents an experimental study of three prevalent power management techniques that are {\em power limitation, frequency limitation}, and {\em ACPI/P-State governor modes} (OS states related to power consumption). Through a benchmark approach with a set of six computing kernels, we investigate {\em power/performance} trade-off with various hardware units and software frameworks (mainly TensorFlow and JAX). Our experimental results show that {\em frequency limitation} is the most effective technique to improve {\em Energy-Delay Product (EDP)}, which is a convolution of energy and running time. We also observe that running at the highest frequency compared to a reduced one could lead to a reduction of factor $\frac{1}{10}$ in EDP. Another noticeable fact is that frequency management shows a consistent behavior with different CPUs, whereas opposite effects sometimes occur between TensorFlow (TF) and JAX with the same power management settings.

Figures (9)

• Figure 1: EDP vs Time on Intel for JAX
• Figure 2: EDP vs Time on Intel for TF
• Figure 3: EDP vs Time on AMD for JAX
• Figure 4: EDP vs Time on AMD for TF
• Figure 5: EDP vs Time on Nvidia for JAX