SIST-TP CEN/CLC/TR 18145:2025

SLOVENSKI STANDARD
01-april-2025
Okoljsko trajnostna umetna inteligenca
Environmentally sustainable Artificial Intelligence
Informationstechnik - Künstliche Intelligenz - Grüne und nachhaltige KI
Ta slovenski standard je istoveten z: CEN/CLC/TR 18145:2025
ICS:
13.020.20 Okoljska ekonomija. Environmental economics.
Trajnostnost Sustainability
35.020 Informacijska tehnika in Information technology (IT) in
tehnologija na splošno general
2003-01.Slovenski inštitut za standardizacijo. Razmnoževanje celote ali delov tega standarda ni dovoljeno.

TECHNICAL REPORT CEN/CLC/TR 18145

RAPPORT TECHNIQUE
TECHNISCHER REPORT
February 2025
ICS 13.020.20; 35.240.01
English version
Environmentally sustainable Artificial Intelligence
Informationstechnik - Künstliche Intelligenz - Grüne
und nachhaltige KI
This Technical Report was approved by CEN on 10 February 2025. It has been drawn up by the Technical Committee
CEN/CLC/JTC 21.
CEN and CENELEC members are the national standards bodies and national electrotechnical committees of Austria, Belgium,
Bulgaria, Croatia, Cyprus, Czech Republic, Denmark, Estonia, Finland, France, Germany, Greece, Hungary, Iceland, Ireland, Italy,
Latvia, Lithuania, Luxembourg, Malta, Netherlands, Norway, Poland, Portugal, Republic of North Macedonia, Romania, Serbia,
Slovakia, Slovenia, Spain, Sweden, Switzerland, Türkiye and United Kingdom.

CEN-CENELEC Management Centre:
Rue de la Science 23, B-1040 Brussels
© 2025 CEN/CENELEC All rights of exploitation in any form and by any means
Ref. No. CEN/CLC/TR 18145:2025 E
reserved worldwide for CEN national Members and for
CENELEC Members.
Contents Page
European foreword . 3
1 Scope . 4
2 Normative references . 4
3 Terms, definitions and abbreviated terms . 5
3.1 Terms and definitions . 5
3.1.1 Terms related to Artificial intelligence and Machine Learning . 5
3.1.2 Terms related to environmental sustainability . 6
3.2 Abbreviated terms . 6
4 Environmental impacts . 7
4.1 General. 7
4.2 A definition of the issue . 7
4.3 European energy infrastructure . 8
5 Energy resources and AI . 10
5.1 General. 10
5.2 Energy consumption by AI . 10
5.3 AI energy efficiency methods . 12
5.3.1 Software solutions . 12
5.3.2 Hardware solutions . 14
5.3.3 Location solutions . 15
6 Non-energy resources and AI . 15
6.1 General. 15
6.2 AI Water use and pollution . 15
6.3 Resources for hardware . 16
6.4 Land use and soil pollution . 16
6.5 Guidance on mitigation solutions . 16
6.5.1 Creating awareness . 16
6.5.2 Considering the complete lifecycle . 17
6.5.3 Circular economy . 17
6.6 Policymaking for mitigation of AI environmental impact . 17
7 Environmental benefits of using AI . 18
7.1 General. 18
7.2 Use cases of energy savings . 18
7.2.1 General. 18
7.2.2 Optimizing the use of natural resources . 18
7.2.3 Helping to manage climate change . 18
7.2.4 AI and sustainable networks in ICT . 19
7.3 Use cases that save other resources . 19
8 Quantification of AI environmental sustainability . 19
8.1 General. 19
8.2 AI carbon footprint estimation . 21
8.3 Comparative measures of efficiency . 24
Bibliography . 26
European foreword
This document (CEN/CLC/TR 18145:2025) has been prepared by Technical Committee CEN/CLC/JTC 21
“Artificial Intelligence”, the secretariat of which is held by DS.
Attention is drawn to the possibility that some of the elements of this document may be the subject of
patent rights. CEN shall not be held responsible for identifying any or all such patent rights.
Any feedback and questions on this document should be directed to the users’ national standards body.
A complete listing of these bodies can be found on the CEN website.
1 Scope
This document provides a description of the main environmental sustainability issues that organisations
or individuals that are developing and/or using Artificial Intelligence (AI) consider, in particular, in the
context of the European energy systems and resources.
It is important to have a focus where AI helps in optimization and virtual deployment of engineering
solutions [1], especially in Europe with limited natural resources. This document reviews the European
AI landscape, with a context of environmental sustainability. This is addressed with a focus on European-
specific aspects of AI demands for resources, as well as its potential to contribute to environmental
sustainability in Europe [2]. The document creates an inventory of impacts and techniques to support
environmentally sustainable use of AI, and an equitable access to computation resources.
Suggested improvements in AI resource management are focused on:
• reduction of the operational AI energy consumption (see section 5)
• reduction of other AI resource consumption (water, etc.) (see section 6)
The document also considers the potential benefits of using AI from a sustainability perspective. Methods
of measuring the environmental sustainability impacts of AI are also quantified.
This document is intended to help with the development of new standards and complement existing
European standards and standardization deliverables that define resource measurement for the use of
AI. It describes best practices and indicates which techniques and management processes for
improvement of AI resource performance and environmental viability. The document is expected to
contribute to voluntary corporate social responsibility (CSR) in Europe, and increase sustainability
awareness for individuals when designing, developing, and using AI. The aim is to create a focus on the
responsible use of AI that prioritizes ethical considerations, human values, and an understanding of the
social implications of AI design and use.
The document is aligned with equivalent activities in ISO/IEC/JTC 1/SC42/WG4, TR 20226 “Green and
Sustainable AI”, but takes into account specific aspects of the European energy system that are not
applicable elsewhere. In particular, sustainable energy supply provided via the European interconnectors
will be taken into account when assessing AI carbon footprint. Additionally AI solutions for the
optimization of energy use will be reviewed and quantified to balance the energy use of AI applications
and services which make extensive use of energy. This report also identifies and addresses the United
Nations Sustainable Development Goals [3, 4]. Additionally, this document aligns with ISO/IEC DIS 21031
Information Technology – Software Carbon Intensity (SCI) [5], ISO/DIS 59004 Circular Economy –
Terminology, Principles and Guidance for Implementation, and the Greenhouse Gas Protocol (GHG),
Product Life Cycle Accounting and Reporting Standard [6].
The upcoming EU AI Act in its current draft encourages voluntary assessment of companies for
environmental sustainability.
2 Normative references
There are no normative references in this document.
3 Terms, definitions and abbreviated terms
For the purposes of this document, the following terms and definitions apply.
ISO and IEC maintain terminology databases for use in standardization at the following addresses:
— ISO Online browsing platform: available at https://www.iso.org/obp/
— IEC Electropedia: available at https://www.electropedia.org/
3.1 Terms and definitions
3.1.1 Terms related to Artificial intelligence and Machine Learning
3.1.1.1
artificial intelligence
AI
research and development of mechanisms and applications of AI systems
Note 1 to entry: Research and development can take place across a number of fields such as computer science, data
science, humanities, mathematics and natural sciences.
[SOURCE: ISO/IEC 22989:2022]
3.1.1.2
artificial intelligence system
AI system
engineered system that generates outputs such as content, forecasts, recommendations or decisions for
a given set of human-defined objectives
Note 1 to entry: The engineered system can use various techniques and approaches related artificial intelligence to
develop a model to represent data, knowledge, process, etc which can be used to conduct tasks.
Note 2 to entry: AI systems are designed to operate with varying levels of automation.
[SOURCE: ISO/IEC 22989:2022]
3.1.1.3
dark data
information assets that organizations collect, process and store during regular business activities, but fail
to use for purposes beyond those associated with the initial collection
3.1.1.4
internet of things
IoT
infrastructure of interconnected entities, people, systems, and information resources together with
services which processes and reacts to information from the physical world and virtual world
[SOURCE: ISO/IEC 20924:2021]
3.1.1.5
machine learning
ML
process of optimizing model parameters through computational techniques, such that the model's
behaviour reflects the data or experience
[SOURCE: ISO/IEC 22989:2022]
3.1.2 Terms related to environmental sustainability
3.1.2.1
environmental sustainability
state in which the ecosystem and its functions are maintained for the present and future generation
[SOURCE: ISO 17889-1:2021]
3.1.2.2
life cycle
evolution of a system, product, service, project or other human-made entity from conception through
retirement
[SOURCE: ISO/IEC/IEEE 15288:2015: Systems and software engineering — Software life cycle processes
Note 1 to entry: ISO/IEC 24748-1:2018 Systems and software engineering — Life cycle management — Part 1:
Guide for life cycle management and ISO/IEC/IEEE 15288:2015 Systems and software engineering — System life
cycle processes provide unified and consolidated guidance of life cycle management of systems and software
3.1.2.3
life cycle assessment
LCA
systematic evaluation of the environmental impact of a product(s) that includes all stages of its life cycle
[SOURCE: ISO 17889-2:2023]
3.1.2.4
circular economy
economic system that uses a systemic approach to maintain a circular flow of resources, by recovering,
retaining or adding to their value, while contributing to sustainable development
Note 1 to entry: The inflow of virgin resources is kept as low as possible, and the circular flow of resources is kept
as closed as possible to minimize waste, losses and releases from the economic system. Resources can be considered
concerning both stocks and flows
[SOURCE: ISO/DIS 59004]
3.2 Abbreviated terms
AI Artificial Intelligence
CPU Central Processing Unit
CSR Corporate Social Responsibility
CNN Convolutional Neural Network
DNN Deep Neural Network
ENTSO-E European Network of Transmission System Operators for Electricity
FFT Fast Fourier Transform
FL Federated Learning
GEMM General Matrix Multiplication
GPU Graphic Processing Unit
GWh Giga-Watt hour
GUM Guide to the Expression of Uncertainty in Measurement
HDC Hyperdimensional Computing
HPC High-Performance Computing
ICT Information and Communication Technologies
IoT Internet of Things
kWh Kilowatt-hour
LCA Life-Cycle Assessment
LLM Large Language Model
ML Machine Learning
NLP Natural Language Processing
NN Neural Network
PUE Power Usage Effectiveness
PV Photovoltaic
RAM Random Access Memory
TSO Transmission System Operator
4 Environmental impacts
4.1 General
The primary environmental impact of AI (and other computer software) is the use of electricity, which is
produced in power stations using various fuels and a variety of power suppliers. Most of these resources
used by AI are from fossil fuels, such as natural gas, oil and coal.
4.2 A definition of the issue
New technologies help improve lives of European citizens and bring major benefits to society and
economy through better healthcare, more efficient public administration, safer transport, image analysis,
optimization of use of resources, a more competitive industry and sustainable agriculture.
AI is developing rapidly and requires nonlinear growth of resources such as electricity and cooling
systems consuming freshwater resources to stabilize the operational use and distributed use of resources
to serve an increasing number of cloud service users.
However, some uses of AI lead to unprecedented energy and resource impact in its life cycles, i.e. training,
deployment, and operational use.
In some European countries, such as France, majority of energy generation is nominated by nuclear fuel.
In other European countries the fuel used is more complex. With the deployment of renewable energy
sources, the complexity and intermittency of multiple types of generation are considered. The fuel mix
used for energy generation in Europe is particularly complex. National Grid operators currently report
fuel mix with a 30-min temporal resolution. The impact of cross-boundary European interconnectors,
which deliver electricity from non-European countries with different fuel mixes, also needs to be
considered. These complexities are analysed along with the associated uncertainties on fuel types to
obtain a combined uncertainty of the operational indirect carbon emissions from the power supply used
by AI. ISO/IEC Guide 98-3 – Uncertainty of Measurement – Part 3 Guide to the expression of uncertainty
in measurement [7, 8] establishes the principles and rules which are used for a broad spectrum of GUM
measurements, not just energy generation.
With broader deployment of renewable energy generation capabilities there is increasing demand for the
allocation of land and construction of the related infrastructure. Photovoltaic generation requires
substantial land mass, wind turbines are installed in offshore waters as well as in mountainous areas and
less common installations. For example, Nevada’s Ivanpah Solar Power Facility uses fields with focussing
mirrors for concentrated solar power, these can be deployed in large exclusion zones, such as deserts.
With the recent disruption of energy supplies seen in 2022, natural gas is sourced from various countries
(not necessarily in Europe), including deliveries of liquified gas by sea. Emissions from ships are direct
emissions alongside other Scope 1 transport emissions. The life cycle of all Scope 1 emissions, those
produced directly by the use of fossil fuel transport systems and energy generation, need to be included
and assessed for the estimation of carbon footprint of AI.
It is important that the life cycle assessment of hardware is included in the assessment of the carbon
footprint of AI. This will involve tracing the supply of rare earth metals, which are not currently mined in
Europe. The hardware emissions are likely to be negligible compared with the operational carbon
footprint of a large language model. It is important to note that many aspects of the life cycle assessment
often remain non-transparent. It is important that these transparency issues are noted in any emissions
assessed.
The focus of the emissions assessment on the impact of the AI corresponds to the Scope 2 indirect
emissions from operational use. These can be immediately quantified using publicly available data
supplied from European national grids. The AI carbon emissions footprint needs to be quantifiable and is
likely to be high impact to understand how AI deployment can be measured in a sustainable way. This is
the recommended approach to measure AI deployments and can be considered responsible AI.
4.3 European energy infrastructure
There are European specific aspects of energy use for AI and there are gaps in current standardization
activities which do not consider the specific energy supply. There are issues for, hydro and nuclear
generation and integrated energy systems (interconnectors). These need to be considered alongside the
United Nations sustainable development goals internationally, in Europe issues related to the delivery of
the European Green Deal, and nationally such as the UK’s Net-Zero-Carbon targets. These specific issues
introduce challenges for the assessment of carbon footprint of energy generation in quantifying the
energy mix used for energy delivery at the point of consumption, for example, at a data centre [9].
Figure 1 — Trans-European energy network ENTSO-E with major electrical interconnectors
International electrical connections are marked by blue lines, and international gas pipes by red lines,
national and minor links are denoted by grey lines. Major interconnectors are critical to understand the
cross-boundary interconnectors for energy in Europe, and the energy mix is key to assessing the energy
carbon footprint of AI use.
No other continent has such a dense interconnected network of high complex energy supply that needs
to be modelled appropriately to estimate the carbon footprint of AI energy use, specifically because
energy is imported from outside of Europe and uses different energy sources. The European Network of
Transmission System Operators for Electricity (ENTSO-E) is the association of the European transmission
system operators (TSOs) representing 35 countries. The synchronous grid of Europe is the largest
synchronous electrical grid (by connected power) in the world. It is an interconnected single phase-
locked 50 Hz mains frequency electricity grid that supplies over 500 million customers, most of the
European Union. The synchronous grid goals are sustainability, affordability and resilience, with a large
number of renewables integrated into many national grids [10].
Eastern European countries rely heavily on brown coal (containing lignite rather than anthracite in black
coal) for their energy supply. In France up to 75 % of electricity is generated by nuclear energy. Carbon
emissions differ depending on the fuel used and the life cycle of energy used. The energy landscape in
Europe is constantly changing, with varying suppliers, domestic and foreign, whose operational
emissions are quantified based on the publicly available data in each European country. In the UK, the
National Grid is providing 30-min fuel mix data, including interconnector data with Scandinavia, Ireland,
and France [11].
AI will be instrumental in improving energy efficiency in the European Network of Trans-European
Energy Network, optimizing energy use in demand-side response programmes, and supporting digital
energy services.
5 Energy resources and AI
5.1 General
Energy is key in the use of computer technologies, and this is particularly the case of the AI systems, as
training and deployment of neural networks can be very energy intensive. The research community is
addressing the emerging challenge of the environmental sustainability of AI with several collaboration
initiatives [12].
AI algorithms that are currently solely evaluated in terms of accuracy, in future will be judged on the
environmental impact of their complete life cycle [13, 14]. An AI system will consume energy at all stages
of its life cycle. Emissions will be created from inception, design and development, verification and
validation, deployment, operation and monitoring, continuous validation and retirement (AI system life
cycle are defined in ISO/IEC 22989:2022). At each step of the AI life cycle, energy resources are
consumed. It is essential that they are quantified, reduced where possible, optimized and efficiently used
to deliver responsible AI.
Where AI systems are using cloud service provision, processing and data handling need to be analysed
with respect to the associated cooling and energy consumed by the corresponding data centres. With
increasing use of multimedia files, in particular, images and videos of high resolution (current photo
matrices reach 45-50 megapixels), data transfer and processing become increasingly higher. Images are
often processed by cloud services without changes to the default device settings, i.e. in full resolution, one
image often reaches 10-15Mb. Contemporary standards of data handling and processing assume
automatic synchronisation of media libraries without much compression by default settings, which most
users never modify. Not only it is necessary to achieve changes to existing industry standards, but it is
also important to raise awareness with users and AI practitioners of good practices for data usage [15]
and application of AI/ML tools.
Citing the German brain researcher and biochemist Henning, “AI is also an energy-guzzling process. The
brain works with 20 watts. This is enough to cover our entire thinking ability. AI needs an incredible
amount of energy to recognize a picture of a penguin from 10 million images. To solve the problem, it
requires entire data centres that need to be kept cool., Because AI aims to reproduce brain activities but
lacks brain’s energy efficiency, we will need a vast number of nuclear power plants to provide the
necessary energy, assuming it would even be possible” [16].
In 2018 OpenAI reported that the amount of compute time used in the largest AI training runs had been
increasing exponentially, doubling every 3.4-month. This is evidenced by an analysis of OpenAI’s compute
time of the largest AI training runs. Additionally, AI is used for targeting customers’ interests online, thus
incentivizing internet browsing and streaming. Large platform companies routinely mine their databases
to address their customers with suggestions for further internet activities. According to the Shift Project
in 2019 much of this increase in compute time has been driven by online video use. This growing use of
compute time is increased by features such as default auto play and smart frame advertisements.
5.2 Energy consumption by AI
There are various side effects of electricity generation due to AI use. For example, direct and indirect
carbon emissions, air and water pollution, disposal of nuclear waste, water recycling for cooling. Amongst
these side effects indirect carbon emissions are currently the biggest social impact in the context of
climate change and require quantification for assessment of environmental sustainability of AI.
In 2015, the Semiconductor Industry Association (SIA) and Semiconductor Research Corporation
published a SIA report which contained a crude estimate of the future energy use required for the Internet
of Things, directly related of the number of raw bit transitions. AI applications were not considered in the
SIA 2015 report though, meaning projections of energy production was underestimated. The report did
not anticipate fast construction of power station that run on coal. In 2022 China built two new coal plants
per week to meet high energy demands. The UK is keeping open the last remaining coal power plant and
Germany has postponed the phase-out of coal as a fuel and has reactivated some coal power plants. The
SIA 2015 report stated that by 2040, digital solutions would exceed the global electric capacity, this was
without the large-scale deployment of AI experienced today in Europe and beyond.
A recent large-scale AI solution, the generative AI model called Evolved Transformer with 110 million
parameters was trained using graphics processing units (GPUs) and consumed about 656 MWh of
electricity which is equivalent to about 284 tonnes of CO emissions, the carbon footprint of a round-trip
transcontinental flight for one person [17]. Later in 2022, a group of Google researchers [18] revised
these estimates and downscaled them, due to the hardware usage and green supply of electricity at
Google. However, such flagship examples provide lower estimates of energy use, which is not average in
industry. Large language models (LLM), such as ChatGPT have require enormous amount of power,
meaning that thousands of processors are needed to both train the models and support the billions of
daily prompts created by users. Each processing unit can consume over 400 W of power while operating.
Typically, LLMs also consume a similar amount of power for cooling and power management as well. This
can lead to up to 10 gigawatt-hour (GWh) power consumption to train a single LLM such as ChatGPT-3
[19]. However, it is necessary to remember that technological development will provide more energy
efficient hardware, and consumption of energy by large models in future be compensated by lower
energy demand. The current report provides the basis for estimation of such impact in general layouts.
The number of AI solutions is growing, and it is essential that the sustainability of AI energy use influences
AI development, not least for the sustainability of the European power supply.
AI requires data, and moving data from its originating point, sensor or information source, consumes
resources. In general, the results of AI processing are smaller in size than the raw data used for the
process. It is recommended that AI processing is carried out as close as possible to the data source,
recognizing that in some cases data needs to be consolidated from several sources before it can be
processed. Solutions for processing the data as close as possible to data sources are emerging and already
available as commercial products: For example:
• Machine learning on Internet of Things (IoT) processors enabling the processing directly in
appliances such as an intelligent scale that performs body parameter analysis and learning inside the
appliance, or rotating machine vibration analysis, that performs machine learning on inexpensive
(€1) processors.
• Federated or collaborative learning where sub-models are created close to the data and consolidated
centrally by transferring the sub-models rather than the data.
• Data compression techniques that use processing to avoid precision loss when targeting artificial
intelligence processes.
Another common resource consumption factor is sensor power. Typically, IoT devices have three stages:
sleep, sense, and connect. With AI at the edge [20], the “sensing” function is associated with a sense-
making function (pattern recognition, classification, elementary transformation step from data towards
knowledge), and this additional AI-related processing will consume energy which needs to be assessed
when choosing and comparing AI strategies, for example, Edge-AI vs Cloud-AI, or Hybrid AI [21]. It is
important to consider that some commercially available edge devices combine on a single VLSI chip a
sensing layer and a pretrained neural network processing layer and therefore are considered AI at the
Edge.
The use of the Internet in general has a large carbon footprint. Electricity consumption from data centres
is 1 % of all energy demand [22], with carbon footprint ranging from 28 to 63 gCO2eq/Gb [23].
Computational load is rarely performed on a stand-alone computer, but rather operates on distributed
compute, high performance computing clusters and cloud service provision [9, 24]. This includes AI
applications, which are often resource demanding. As was shown in [17], training a single natural
language processing model can emit 284,000kg of CO , equivalent to 125 round-trip flights between New
York and Beijing.
Emissions that are produced directly by combustion of fossil fuels are Scope 1 emissions. These emissions
are created using fossil fuel transport systems and energy generation. AI will help reduce Scope 1
emissions via smart interventions (demand-side response, optimization of combustion, etc.) Scope 2 are
indirect emissions from electricity use, and AI can play a major role in reducing these emissions. Scope 3
emissions are produced during the life cycle use of technology, these emissions are important in the
assessment of emissions which would be created by an AI solution and are in the scope of this report.
Scope 4 Emissions are avoided emissions. AI has great potential for quantifying avoided emissions and
the resultant carbon savings, these are addressed in this report.
5.3 AI energy efficiency methods
5.3.1 Software solutions
The following software-based methods can help improve the carbon footprint of AI:
— Use of pre-trained modules and optimization (re-use, transfer of training algorithms)
— Use of one-for-all neural networks
— Federated learning
— Compression algorithms
— Frugal AI (using smaller data sets to train models)
— Avoidance of dark data, which is sometimes stored by organisations beyond the term of use
Software-based optimization of resources is achieved in several ways, for example by using pre-trained
modules, parallelisation, and optimal algorithms. The AI community has started to recognize their
responsibility to develop sustainable AI solutions which reduce computational load [25], such as Once-
for-all neural networks [26]. One method to address the reduction of the volume of data transfer and the
associated resource cost is federated learning (FL). This approach implements a distributed machine
learning process in which different parties collaborate to jointly train a machine learning model without
the need to share training data with the other parties [27]. This approach also has the advantage of
facilitating the application of different regulatory and privacy requirements that would apply on data
transfer, by only transferring models that do not contain such regulated information.
The large number of parameters in most of the state-of-the-art deep neural networks (DNN) make them
compute intensive, resulting in high infrastructure demands and latency [28]. Hence, it is crucial to
reduce the computational demands of DNNs to successfully meet the latency and resource requirements
of a production environment (which can be a shared resource like cloud, or a low resource system like
mobile phones). Common state-of-the-art compression mechanisms for convolutional neural networks
(CNN) include filter and weight pruning, sparsification, quantization, binarization [29], encoding, low
rank decompositions, thresholding, etc. More advanced compression techniques for BERT use
distillation, word vector elimination, and others [30, 31].
New solutions are being created to avoid training a model each time a combination of use-case or device
changes [25, 26]. These solutions design a once-for-all network that can then be configured in accordance
with the target environments and use-cases, sharing the training cost between all of the target
environments [32].
Building algorithms often consumes more resources than running them. Therefore, it makes sense to
reuse algorithms as much as possible [31]. To make more energy efficient AI systems, companies and
universities work together on an open-source basis [33] because building an AI service from scratch uses
the most data and is therefore the most energy intensive. (For more information see [34])
Energy efficient approaches are leading to the creation of Edge AI solutions, these are implemented in
some commercially available single chip systems with a sensor layer and a pretrained neural network
layer for AI processing.
During the learning phase of AI models, a large amount of data are required to achieve the best precision
possible (without over training). This is the case for example with Natural Language Programming (NLP)
where many data sets are available online, although sometimes not free and not labelled, indicating an
inequality in data accessibility. For example, in some domains edge computing where data are scarce
finding energy efficient solutions to train models could help reduce the need for data and eventually
create solutions which require reduced energy consumption.
Training of an AI model with smaller data sets which achieve precision is then transferable to algorithms
which currently require large training data sets. This is a field of research known as Frugal AI [35]. One
approach is to create a “teaching” block see the transfer-learning structure diagram below. A Frugal AI
approach would be to learn on existing models by recycling old algorithms to create new ones. This Frugal
AI approach is also known as transfer learning. The concept of transfer learning comes from the human
ability to apply previous knowledge to new situations. These techniques are already used in NLP and
computer vision, which are two data intensive domains [36].

Figure 2 — Transfer learning structure [35]
In some cases, it is possible to achieve better results using transfer learning [37] because the fine-tuning
of the hyper parameters is better when using models already trained with large data sets on a given
subject.
As energy becomes more valuable, financially and environmentally, reusing existing algorithms for AI
services could bring significant energy efficiency gains [14]. The extraction of structural relations and
features, such as the ones obtained through physical or biological modelling, reduces the need in the
system where artificial intelligence, i.e. machine learning, deep learning, reinforcement learning is still
necessary. An efficient mathematical structure of the AI implementation provides further efficiency gains,
for example Fast Fourier transformation versus classical Fourier transform, operations regularity and
simplification through a butterfly type of computation arrays, etc.
Many digital applications such as IoT sensors create large quantities of data that are stored but are often
of very little use to any application, including AI. This data uses energy when it is generated, transported,
and stored. AI systems need to avoid importing any data that is of no use. Data filtering techniques help
optimize the data that are ingested by AI systems.
Where data are to be kept long-term (for example, for compliance purposes) but is used only rarely, the
storage solutions in low energy environments and formats (such as tape or laser disks) are to be
considered.
In some circumstances using AI is not the optimal solution, or its benefits are outweighed by
environmental impacts. In these circumstances, alternative processing methods are used. This is a central
issue of circular economy and sustainability when considering the trade-offs between AI’s benefits and
negative environmental impacts. The current report provides an approach with a robust methodology
for assessment as whether the environmental negative impacts will be higher than the benefits of AI.
5.3.2 Hardware solutions
The hardware-based methods that can help improve the carbon footprint of AI:
— Reduced precision hardware and hardware accelerators
— in-memory computing methods
— neuromorphic computing
— Edge computing
— New technologies using task distribution, including optic fibre innovations.
Recently reduced precision techniques have proven exceptionally effective in accelerating deep learning
training and inference applications. The most used arithmetic function in deep learning is the dot product,
which is the building block of generalized matrix multiplication (GEMM) and convolution computations.
With current hardware architectures a dot product requires two floating-point computations requiring
accumulation inside processors. Research work verified experimentally the minimum accumulation
precision required for these computations, this is targeted to dramatically cut AI training time and cost.
Research has developed a way to enable 4-bit training of deep learning models that can be used across
many domains. Reduced accumulation bit-width requirement translates directly to a 1,5 –
2,2 × improvement in hardware energy efficiency [38].
The human brain has no separate compartments to store and compute data (“no hard drive, only RAM”),
and therefore consumes significantly less energy. Artificial Intelligence models are typically stored in off-
chip memory, and computational tasks require a constant shuffling of data between the memory and
computing units. This is a process that slows down computation and limits the maximum achievable
energy efficiency. In-memory computing methods using resistance-based memristive storage devices are
showing promising results as are non-von Neumann approaches for developing hardware that can
efficiently support AI inference models [39].
Neuromorphic computing which encompasses devices that can mimic the natural biological structures of
the human nervous system presents a promising energy efficient alternative over conventional
computing architectures [40]. High-performance computing (HDC), as an emerging computational
paradigm aims to mimic attributes of the human brain’s neuronal circuits such as hyper dimensionality,
fully distributed holographic representation, and (pseudo) randomness with the energy efficiency of
human brain [41]. This research work was performed in collaboration with ETH Zürich and was
supported in part by the European Research Council [42].
AI on Edge is low energy processor machine learning. Low energy processors with embedded machine
learning for IoT are technologies for powering devices on the edge. It brings analytics and machine
learning to the on-premises edge environment, enabling machine performance optimization, proactive
maintenance, and operational intelligence at a much lower energy cost and better efficiency. These
solutions address latency (there's no round-trip to a server), privacy (no personal data leaves the device),
connectivity (internet connectivity is not required), size (reduced model and binary size) and efficient
power consumption (efficient inference and a lack of network connections).
The European Commission will support the development of AI on Edge with dedicated testing and
experimentation facilities under the Digital Europe Programme. The Edge AI testing and experimentation
facilities aim, as a European platform, to enable companies of any size to test and experiment innovative
edge AI components based on advanced low-power computing technologies and advanced connectivity
[43, 44]. The Digital Europe Programme will also fund other testing and experimentation facilities in four
sectors (agri-food, healthcare, manufacturing, and smart cities/communities) that in part are focused on
reducing AI’s carbon footprint and/or use AI to reduce the respective sector’s carbon footprint.
Large-scale distributed training of deep neural networks on state-of-the-art platforms is expected to be
severely communication constrained and induce higher energy consumption
...