Untitled

Environmental Impacts of Digital Technologies

Methodology Overview

  • Purpose: Assess the environmental impact of Stable Diffusion as an end-to-end service.

  • Considerations include:

    • Carbon impact.

    • Consumption of metals.

    • Life Cycle Assessment (LCA) approach.

  • Addressing current methods of estimating electricity consumption of training Generative AI (GenAI) models.

  • Proposal for a transparent and reproducible method based on experimental sampling.

  • Differentiation of operational emissions and embodied emissions during hardware life cycle phases.

Current Approach to Environmental Impact in AI

  • Focus: Sustainability of computing technologies, specifically GenAI services like ChatGPT and Stable Diffusion.

  • Importance of evaluating environmental impact to understand sustainability better.

Review of Generative AI

  • Definition: Generative AI focuses on creating content like text and images, first coined in 2014, gaining prominence post-2022.

  • Developing GenAI models requires:

    • Selecting model structure and learning algorithms.

    • Extensive data collection and learning processes.

  • Training phase demands resources, typically utilizing multiple GPUs.

    • Example: Stable Diffusion used 258 GPUs and 64 CPUs for training.

    • Training involves hundreds of thousands of steps.

Environmental Impact Concerns

  • Growing popularity of GenAI services prompts questions of sustainability.

  • Inquiry into the magnitude of environmental impact due to increased electricity demand and IT infrastructure.

Environmental Footprint of AI

  • Studies since 2019 assessing electricity consumption during machine learning training.

  • Common methods to estimate electricity consumption:

    • Thermal Design Power (TDP):

    • Represents max power of components used to estimate energy usage, but limited as it only accounts for certain components.

    • Power Meters (PMs):

    • External PMs: Most accurate but impractical without direct access.

    • Software-based PMs: More precise than TDP but still incomplete since they require measurements during execution.

  • Overall, literature lacks comprehensive methodologies that are accessible and reproducible in assessing GenAI's environmental impact.

Life Cycle Assessment (LCA)

  • Definition: Multi-criteria evaluation based on ISO 14040 and 14044 standards, focusing on potential environmental impacts considering all life cycle phases:

  • Phases of LCA:

    1. Definition of objectives.

    2. Life Cycle Inventory (LCI) analysis.

    3. Life Cycle Impact Assessment (LCIA).

    4. Interpretation of results.

  • Benefits of LCA for GenAI include addressing environmental impacts beyond carbon emissions alone.

Enhanced Tools for Measuring Environmental Impact

  • Expansion of the evaluation framework to include additional impact categories specific to AI:

    • Abiotic Depletion Potential (ADP): Related to the decrease of minerals and metals resources.

    • Global Warming Potential (GWP): Contribution to climate change.

    • Primary Energy (PE): Cumulative energy demand.

  • LCA employed to assess the cost of the average user journey in Generative AI services.

Energy Consumption Assessment

  • Replicated the training process for Stable Diffusion to monitor electricity consumption.

  • Assumptions based on constant energy consumption of epochs, leading to regression models predicting energy consumption tied to training steps.

    • Equations from training:
      E<em>256=5.26e−04N+2.01e−02E<em>{256} = 5.26e^{-04}N + 2.01e^{-02} E</em>512=1.78e−03N+1.64e−02E</em>{512} = 1.78e^{-03}N + 1.64e^{-02}

  • Conducted experiments on nodes from the Grid’5000 platform, utilizing monitoring tools for accurate assessments.

Findings of LCA for Stable Diffusion

  • Estimated significant environmental impact:

    • 463 tons of CO2 equivalent identified through LCA, highlighting substantial implications of digital services.

    • Distribution of impacts showed that terminals and networks represent a significant share of the GenAI service impact.

Recommendations and Implications

  • Need for extending the lifespan of hardware and software to mitigate the embodied emissions related to GenAI services.

  • Emphasis on developing methodologies that enhance transparency, especially regarding electricity consumption and deployment practices.

Conclusion

  • The assessment of GenAI's environmental impact is critical for determining sustainability.

  • Proposed methodologies highlight the importance of comprehensive evaluations that account for multi-category environmental impacts and accurate energy estimations.

  • Encouragement for improvement in data transparency to facilitate better understanding and informed decisions in sustainability practices for GenAI services.