The Hidden Environmental Cost of Poor Coding Practices in TensorFlow and Keras Applications: A Study on Resource Leaks and Carbon Emissions

2026-06-19 · Source: cs.SE updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Software Development & Engineering · Depth: Advanced, extended

Summary

An empirical study quantifies the environmental impact of poor coding practices, specifically resource-leak "smells," in TensorFlow and Keras applications. Researchers investigated two common smells: Improper Model Reuse (IMR) and Unreleased Tensor References (UTR). Controlled experiments, using a CNN trained on CIFAR-10 for 50 epochs on an NVIDIA Tesla T4 GPU, revealed significant increases in energy consumption and CO2 emissions. IMR elevated electricity usage by approximately 31.78% (from a baseline of 0.025646 kWh to 0.033797 kWh) and CO2 emissions by 31.78% (from 0.011608 kg to 0.015297 kg). UTR caused even larger increases, raising electricity consumption by 45.80% (to 0.037393 kWh) and CO2 emissions by 45.77% (to 0.016921 kg). These statistically significant findings, monitored via CodeCarbon, highlight that resource mismanagement measurably degrades ML energy efficiency.

Key takeaway

For MLOps Engineers and ML Scientists deploying TensorFlow/Keras applications, you must prioritize resource lifecycle management. Your seemingly minor coding decisions, like improper model reuse or unreleased tensor references, can increase energy consumption by 32-46% and proportionally raise CO2 emissions. Integrate resource cleanup and energy monitoring into your development and deployment pipelines to mitigate these hidden environmental and operational costs.

Key insights

Resource-leak code smells in ML applications significantly increase energy consumption and CO2 emissions.

Principles

• Resource management is a first-class software quality attribute.
• ML code quality directly impacts environmental sustainability.
• Minor coding decisions can incur measurable environmental costs.

Method

A repeated-measures experiment compared a clean TensorFlow/Keras CNN baseline against independently injected Improper Model Reuse (IMR) and Unreleased Tensor References (UTR) smells. Energy and CO2 were monitored using CodeCarbon across ten runs.

In practice

• Integrate resource cleanup practices into ML development.
• Incorporate energy monitoring into ML workflows.
• Use smell detection tools for ML-specific resource leaks.

Topics

• TensorFlow
• Keras
• Resource Leaks
• Carbon Emissions
• Energy Efficiency
• Code Smells

Best for: AI Engineer, Research Scientist, AI Scientist, Machine Learning Engineer, MLOps Engineer

Related on AIssential

• Officials hugely underestimated impact of AI datacentres on UK carbon emissions - The Guardian — UK government vastly underestimated AI datacentre carbon emissions, raising climate concerns.
• Claude settles the Queries Water Wastage Debate — Local apps consume vastly less water for simple queries than Google search or AI, due to local computation.
• How some data center operators are tackling their water use problems — Data center water use is a complex, localized issue with inherent trade-offs between water consumption and energy efficiency.
• Standardized Methods and Recommendations for Green Federated Learning — A standardized carbon accounting method for Federated Learning is crucial for reproducible "green" FL evaluation.
• Sustainable AI: A Beginners Guide — Sustainable AI minimizes environmental harm by optimizing energy efficiency, renewable energy sourcing, and transparency.
• Data centers and crypto could hike power costs by 57% by 2030 — Data center and cryptocurrency expansion threatens to reverse U.S. power sector gains, driving up costs and CO2 emissions by 2030.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.SE updates on arXiv.org.