This guide helps you understand how everyday coding, data use, or AI tools contribute to environmental impacts, and what you can realistically do to measure such impacts. It offers practical starting points: how to measure, reflect, and improve. Whether you're a researcher, developer, or policymaker, it helps you make more informed and responsible decisions.
What is green computing?
According to the Brundtland Commission (1987) sustainability is defined as “meeting the needs of the present without compromising the ability of future generations to meet their own needs.” There are several definitions of environmentally sustainable (also called “green”) computational workflows (see e.g. the reference below at [1], or the definition at this link or this link). We follow the definition of sustainability above, and narrow it down to (1) Activities concerning computing, and to (2) preserving the environment as one of the enablers to meet future generations’ needs.
In this guide, computing encompasses every aspect and resource that is necessary for a computational workflow. Broadly speaking, this includes the hardware, software, and data that enable these workflows.
These elements have an impact on the environment, including (but not limited to) the energy cost to power hardware when running software workflows or storing and moving data, the impact on natural resources of mining the material needed to build the hardware, the water needed both to produce energy and to cool the hardware, as well as the electronic waste created during hardware disposal.
Therefore, we define green computing as practices that minimise the negative impact of computing on the environment.
Why is it necessary?
The environmental impact of computing, in research and beyond, has historically been under-recognised. Users rely on computing hardware and electricity costs that are not immediately visible. For example, in the research sector, organisations generally cover the computing costs (e.g. electricity bills) centrally, creating the impression that computing power is a free resource. Another example is the use of generative AI tools (e.g. large language models) or software-as-a-service: the user pays a subscription for the computing power running in a data centre, but does not directly see the energy consumption and impacts related to the data centre.
However, computing has a significant impact in terms of environmental resources. Examples of these impacts are below:
- The global electricity consumption of data centres is a driver of greenhouse gas (GHG) emissions, and in particular, CO2 emissions. It is estimated at 126 MTCO2e for 2020 [6]. While emissions depend on how electricity is sourced [source], electricity consumption from data centres is projected to at least double from 2024 to 2030 and could reach 15% of the global electricity consumption [source][source][source]. In particular, an affordable, reliable and sustainable electricity supply will be necessary for further developments of AI, as a typical AI data centre consumes as much electricity as 100000 households [source].
- Computing systems, particularly data centres and AI infrastructure, can consume significant amounts of water, directly for cooling, and indirectly, for electricity generation [2].
- Data centres are localised consumers of electricity and water, as well as CO2 and GHG emitters. This aspect can be exacerbated by the presence of generators required for continuous uptime or consumption spikes. This puts pressure on the community, infrastructure and ecosystem where they are located, and creates challenges for local integration.
- Computing hardware relies on the extraction of rare earth elements and critical minerals, whose mining processes can cause significant environmental damage, including large-scale waste generation, land disruption, and pollution [source].
- Electronic waste (e-waste) is one of the fastest-growing waste streams globally and contains hazardous materials that can harm ecosystems and human health. E-waste is growing rapidly and is projected to reach 82 million tonnes by 2030 [source].
The main goal of this introduction is to raise awareness of the environmental impact of computing. Because this impact and its costs are diffuse and not borne by the individual, it is not directly visible. Quantifying and surfacing this impact enables more informed and responsible decision-making in the use, design and execution of computational workflows. Moreover, the impact of an overall mindset [3][4] and policy change that requires computing providers to be transparent on their environmental impact, and subsequently mitigate it, can be sizable.
Furthermore, this guide provides pointers on how to make one’s own computational workflow more environmentally sustainable, adding an environmental cost dimension to computational workflow optimisation. Implementing and quantifying such improvements not only leads to an improved understanding of the environmental consequences of computational choices, but also motivates broader adoption of sustainable practices, including seeking out, developing and applying technological improvements that can be implemented at scale. Just to name a few examples: work from the University of Glasgow [5] studying more power-efficient computing chips led to the implementation in large-scale, high-throughput scientific computing infrastructures [6]; a case study between Finnish and African researchers shows that a redesign of their software architecture reduced energy consumption by more than 60% [link]; a study performed by AI researchers from Carnegie Mellon University and Hugging Face reports on modelling of the optimisation of large language models in real-world settings as a prerequisite for implementation [7][A]; finally, shifting the execution of computing workflows to different times, geographical location or compute resource can reduce the energy demands [B].
It should also be noted that improved energy efficiency can drive increased demand, as per Jevons paradox, which is a rebound effect [10][11]. This can be summarised as the phenomenon where, if a resource becomes more efficient, then it also becomes cheaper, leading to more extensive use. This effect is particularly noticeable in computing workflows involving large AI models, whose demand keeps growing as energy efficiency is improved [12]. Other such examples exist, and qualitative and quantitative research is ongoing to inform effective mitigation techniques [13]. Nevertheless, understanding why and how to improve environmental sustainability aspects is still a crucial factor in the overall picture.
Software sustainability, understood through the FAIR4RS [14] principles and advocated by the SSI, is fundamental to individual and collective green computing strategies: building reproducible, well-documented and overall better code encourages a more thought-through approach to the consequences of computing workflows on others.
Measuring the environmental impact of a computational workflow: a quickstart guide
An essential first step in reducing the environmental impact of computational workflows is measuring it. This information is useful to workflow developers and users, with examples of self-reporting in [15], and it is increasingly becoming part of recommendations to policy makers and funders (see the examples of the UK [16] and France [17]).
However, measuring the entire environmental impact of a computational workflow is complex. This process encompasses hardware production, transport, use and disposal, and mirrors how we might assess the environmental cost of a can of tomato sauce: natural resources are impacted by growing the tomatoes, processing, packaging, transport to the supermarket, and waste management. An accurate measurement of all these components requires a full Lifecycle Assessment (LCA) following international standards.
To make this task conceptually more manageable, we split the environmental impacts of a computational workflow into two categories: embodied impacts, related to the manufacturing and disposal of the computing hardware, and operational impacts, stemming from the energy consumed to run the computational workflow. Given this complexity, it is therefore important to define the boundaries of a computational workflow measurement, as suggested by the Green Software Foundation’s Software Carbon Intensity (SCI) (ISO/IEC 21031:2024), which defines a methodology to calculate the rate of carbon emissions for a software system. This approach involves several steps: first, deciding on what to include in the measurement; second, choosing a standardised measure of what the component is doing (e.g. inference for AI systems, or reading a file for a data analysis software) to understand how the environmental impacts will scale; third, define a measurement technique for each of these components and units, and then finally measuring and reporting the results and the chosen methodologies.
While this introduction does not give a complete list of methods and tools to perform a lifecycle assessment or calculate the SCI, it can serve as a quick-start guide with resources that can be used as a first step for individual computational workflows, in terms of:
- Embodied impacts: a full lifecycle assessment for the hardware, which often needs to be done by the hardware manufacturers (see e.g. this example). Relevant information is starting to be collected in databases such as the Boavizta database [18]
- Operational impacts: An operational metric for the environmental impact of computational workflows is their energy consumption during execution, which can then be transformed into other indicators of environmental impact, such as CO2 emissions. This can, for example, be estimated either with measurements while the software runs (see [19, 20] for a list of tools that can be used by software practitioners) or through modelling using online calculators [21]. It is important to note that measuring and reducing energy consumption (including the optimisation of idle times, where the computing servers still consume energy) is in active development for High Performance Computing, see for example [22].
Building on these measurement approaches, the reader can then go on to monitor and moderate the environmental impact of their software workflow. In the guides following this introduction, the reader will find further information and checklists on good practices for green data, code efficiency assessment, and writing energy-efficient code on appropriate hardware.
Takeaway message and some next steps
Green computing doesn’t require perfection, as it can start with awareness and small, intentional changes. Begin by measuring your current impact, then look for simple optimisations, like reducing unnecessary runs or improving efficiency. Over time, these small steps can scale into meaningful change. As a next step, you can try one of the tools below to estimate your footprint and explore ways to integrate sustainability into your regular workflow. Finally, the other SSI green computing guides highlight further practical steps on how to minimise the operational impacts on the environment of your computing workflows.
Useful resources:
Here are some tools useful for estimating the carbon footprint of a computational workflow (also included in [19, 20]:
- Tools to use if you want to run the code and see the impact:
- CodeCarbon: https://codecarbon.io/
- Scaphandre: https://github.com/hubblo-org/scaphandre
- CarbonTracker: https://carbontracker.info/
- Tools to use if you see the impact without running the code. These do a post-hoc analysis based on models, without execution:
- Green Algorithms calculator: https://calculator.green-algorithms.org/
- ML CO₂ Impact calculator: https://calculator.linkeddata.es/
Declaration of Delegation to Generative AI (GAIDeT)
The authors declare the use of generative AI in the research and writing process. According to the GAIDeT taxonomy (2025), the following tasks were delegated to GAI tools under full human supervision: Proofreading and editing, Formulation of conclusions, Citation formatting.
The GAI tools used were: Claude Haiku 4.5, Sonnet 4.5, ChatGPT 4.5.
Responsibility for the final manuscript lies entirely with the authors. GAI tools are not listed as authors and do not bear responsibility for the final outcomes.
Declaration submitted by: Jyoti Bhogal, Caterina Doglioni
Additional note: We used Claude models to help with suggestions of transitions between paragraphs to increase coherence, and ChatGPT to generate an initial draft of the introduction and take-away messages. We recognise the environmental impact of the use of GenAI for text editing, as well as the difficulty in obtaining an estimate that is relevant for this guide. As a first step, based on this link and this link, we estimate the energy use involved in queries concerning this guide between 3 Wh (using 10 “average queries” as lower limit) and 400 Wh (10 x Claude code sessions as upper limit), where the system boundaries cover exclusively the operational costs of model queries (inference).
Citations:
[1] S. Murugesan. 2008. Harnessing Green IT: Principles and Practices. IT Professional 10, 1 (January-February 2008), 24–33. https://doi.org/10.1109/MITP.2008.10
[2] David Mytton. 2021. Data centre water consumption. npj Clean Water 4, 11 (2021). https://doi.org/10.1038/s41545-021-00101-w
[3] Heike Rau, Stephan Nicolai, and Silja Stoll-Kleemann. 2022. A systematic review to assess the evidence-based effectiveness, content, and success factors of behavior change interventions for enhancing pro-environmental behavior in individuals. Frontiers in Psychology 13 (2022), 901927. https://doi.org/10.3389/fpsyg.2022.901927
[4] Hunt Allcott. 2011. Social norms and energy conservation. Journal of Public Economics 95, 9-10 (2011), 1082–1095. https://doi.org/10.1016/j.jpubeco.2011.03.003
[5] Emanuele Simili, Gordon Stewart, Samuel Skipsey, Dwayne Spiteri, Albert Borbely, and David Britton. 2024. ARMing HEP for the future Energy Efficiency of WLCG sites (ARM vs. x86). EPJ Web of Conferences 295 (2024), 11007. https://doi.org/10.1051/epjconf/202429511007
[6] Jens Malmodin, Nina Lövehagen, Pernilla Bergmark, and Dag Lundén. 2024. ICT sector electricity consumption and greenhouse gas emissions – 2020 outcome. Telecommunications Policy 48, 3 (2024), 102701. https://doi.org/10.1016/j.telpol.2023.102701
[7] Jillian Fernandez, Chaitanya Na, Vivek Tiwari, Yonatan Bisk, Sasha Luccioni, and Emma Strubell. 2025. Energy considerations of large language model inference and efficiency optimizations. In Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), 32556–32569
[8] Felipe Oviedo, Fiodar Kazhamiaka, Esha Choukse, Allen Kim, Amy Luers, Melanie Nakagawa, Ricardo Bianchini, and Juan M. Lavista Ferres. 2025. Energy Use of AI Inference: Efficiency Pathways and Test-Time Compute. arXiv:2509.20241 [cs.LG]. https://arxiv.org/abs/2509.20241
[9] Nicolas Tirel, Philippe Roose, Sergio Ilarri, Adel Noureddine, and Olivier Le Goaër published "Workload Shifting Techniques: From Digital Inebriation to Sobriety" in ACM Computing Surveys 58(5), 2025 (pp. 1–36). https://doi.org/10.1145/3769301 .
[10] Blake Alcott. 2005. Jevons' paradox. Ecological Economics 54, 1 (2005), 9–21. https://doi.org/10.1016/j.ecolecon.2005.03.020
[11] Steve Sorrell. 2009. Jevons' Paradox revisited: The evidence for backfire from improved energy efficiency. Energy Policy 37, 4 (2009), 1456–1469. https://doi.org/10.1016/j.enpol.2008.12.003
[12] Dongyang Yu and Bingjie Xu. 2026. The Jevons Paradox in the AI era: Artificial intelligence adoption for enhancing environmental sustainability at the firm level. Economic Analysis and Policy 90 (2026), 946–966. https://doi.org/10.1016/j.eap.2026.01.060
[13] Alexandra Sasha Luccioni, Emma Strubell, and Kate Crawford. 2025. From Efficiency Gains to Rebound Effects: The Problem of Jevons' Paradox in AI's Polarized Environmental Debate. In Proceedings of the 2025 ACM Conference on Fairness, Accountability, and Transparency (FAccT '25). Association for Computing Machinery, New York, NY, USA, 76–88. https://doi.org/10.1145/3715275.3732007
[14] Michelle Barker, Neil P. Chue Hong, David S. Katz, et al. 2022. Introducing the FAIR Principles for research software. Scientific Data 9 (2022), 622. https://doi.org/10.1038/s41597-022-01710-x
[15] Qin, Y., Havulinna, A.S., Liu, Y., et al. 2022. Combined effects of host genetics and diet on human gut microbiota and incident disease in a single population cohort. Nature Genetics 54 (2022), 134–142. https://doi.org/10.1038/s41588-021-00991-z
[16] Martin Juckes, Michael Bane, Jennifer Bulpett, Katie Cartmell, Miranda MacFarlane, Molly MacRae, Alex Owen, Charlotte Pascoe, and Poppy Townsend. 2023. Sustainability in Digital Research Infrastructure: UKRI Net Zero DRI Scoping Project final technical report. Zenodo, August 1, 2023. https://doi.org/10.5281/zenodo.8199984
[17] How to include environmental sustainability criteria in national AI funding schemes? Reflecting on the example of France and the Green Algorithms tool. Zenodo, January 7, 2025. https://doi.org/10.5281/zenodo.14607021
[18] Thibault Simon, David Ekchajzer, Adrien Berthelot, Eric Fourboul, Samuel Rince, and Romain Rouvoy. 2025. BoaviztAPI: A Bottom-Up Model to Assess the Environmental Impacts of Cloud Services. SIGENERGY Energy Informatics Review 4, 5 (December 2024), 84–90. https://doi.org/10.1145/3727200.3727213
[19] Rene Schiffmann. 2025. UofM-Green-Compute/Energy-Evaluation-Tools: First release of Green Computing Tools Table (v1.0). Zenodo, 2025. https://doi.org/10.5281/zenodo.16780717
[20] Boavizta. ICT Sustainability Tools. Retrieved March 31, 2026, from https://boavizta.github.io/ict-sustainability-tools/
[21] Loïc Lannelongue, Jason Grealey, and Michael Inouye. 2021. Green Algorithms: Quantifying the Carbon Footprint of Computation. Advanced Science 8, 12 (June 2021), 2100707. https://doi.org/10.1002/advs.202100707
[22] Czarnul, Pawel, Proficz, Jerzy, Krzywaniak, Adam, Energy-Aware High-Performance Computing: Survey of State-of-the-Art Tools, Techniques, and Environments, Scientific Programming, 2019, 8348791, 19 pages, 2019. https://doi.org/10.1155/2019/8348791
Acknowledgements
This guide by written by Jyoti Bhogal Independent Researcher & SSI Fellow, and Caterina Doglioni, University of Manchester. The guide was reviewed by Löic Lannelongue and Diego Alonso Alvarez.