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

2025-05-06

AI Summary

This study systematically investigates power management mechanisms and energy-efficiency–performance trade-offs of JAX and TensorFlow on CPU/GPU platforms. We evaluate energy consumption and latency across six computational kernel benchmarks under three hardware-level power control strategies: static power capping, dynamic frequency scaling, and ACPI/P-State governor policies. Our key findings are threefold: (1) Frequency limiting yields the most substantial improvement in Energy-Delay Product (EDP)—operating at maximum frequency reduces EDP by up to 90% (i.e., 1/10), with an average 10× improvement consistent across diverse CPU architectures; (2) JAX and TensorFlow exhibit opposing energy-efficiency trends under identical power constraints, revealing a previously unrecognized deep coupling between framework-level scheduling and hardware power management; (3) These results provide empirical foundations and a novel co-design paradigm for software–hardware collaborative optimization targeting AI training and inference workloads.

Technology Category

• Machine Learning: Hardware-aware ML
• Humans and AI: Other Foundations of Human Computation & AI
• Computer Vision: Learning & Optimization for CV

Application Category

• Systems and Infrastructure for Web, Mobile and WoT: Energy management for devices in mobile Web and WoT environments
• Graph Algorithms and Modeling for the Web: Efficient manipulation of static and dynamic Web-related graphs
• Economics, Online Markets and Human Computation: Cost models of using LLMs in production systems

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.

Problem

• Study power management techniques in CPU/GPU computing
• Compare power/performance trade-offs using TensorFlow and JAX
• Evaluate effectiveness of frequency limitation on energy efficiency

Innovation

• Benchmarking power management techniques with JAX and TensorFlow
• Frequency limitation improves Energy-Delay Product (EDP)
• Comparing CPU/GPU power behavior across frameworks