A Systematic Review of Green AI: Roberto Verdecchia June Sallou Luís Cruz

A Systematic Review of Green AI
Roberto Verdecchia June Sallou Luís Cruz

University of Florence TU Delft TU Delft
Florence, Italy Delft, The Netherlands Delft, The Netherlands
[email protected] [email protected] [email protected]
Abstract A Systematic Review of Green AI
With the ever-growing adoption of AI-based systems, the carbon Roberto Verdecchia, June Sallou, Luis Cruz
footprint of AI is no longer negligible. AI researchers and practi- Systematic literature Type of studies
review
tioners are therefore urged to hold themselves accountable for the
98
arXiv:2301.11047v3 [cs.AI] 5 May 2023
carbon emissions of the AI models they design and use. This led in primary
recent years to the appearance of researches tackling AI environ- studies Position papers Observational Solution papers
(11) studies (35) (52)
mental sustainability, a field referred to as Green AI. Despite the
Green AI topics
rapid growth of interest in the topic, a comprehensive overview of
Green AI research is to date still missing. To address this gap, in this
paper, we present a systematic review of the Green AI literature. Footprint Hyperparameter Benchmarking Deployment Precision/energy Algorithm Other
From the analysis of 98 primary studies, different patterns emerge. monitoring tuning trade off design
The topic experienced a considerable growth from 2020 onward. Recurrent Green AI Most considered AI Most considered artefacts
domains
Most studies consider monitoring AI model footprint, tuning hy- phase
perparameters to improve model sustainability, or benchmarking

models. A mix of position papers, observational studies, and solu- AI models Data AI pipeline
Training phase
tion papers are present. Most papers focus on the training phase, General Edge computing
Additional findings
are algorithm-agnostic or study neural networks, and use image Recurrent Algorithm
data. Laboratory experiments are the most common research strat- Neural
egy. Reported Green AI energy savings go up to 115%, with savings network Research Energy Industry Tool
over 50% being rather common. Industrial parties are involved in strategies savings Involvement availability
Green AI studies, albeit most target academic readers. Green AI tool

provisioning is scarce. As a conclusion, the Green AI research field
results to have reached a considerable level of maturity. Therefore,
Graphical Abstract: From a systematic review of the Green
from this review emerges that the time is suitable to adopt other
AI literature, Green AI results to focus on solutions, and
Green AI research strategies, and port the numerous promising
is often not bound to a specific context or algorithm. The
academic results to industrial practice.
Green AI research field results to be mature, i.e., the moment
is suitable to port results from academic research to indus-
trial practice.
1 Introduction system’s lifecycle (e.g, data collection, training, monitoring), dif-

In recent years, the Artificial Intelligence (AI) community has been ferent artifacts (e.g., data, model, pipeline, architecture, hardware),
challenged to bring the carbon footprint of AI models to the top etc [3].
of their research agenda. The iconic paper by Strubell et al. [97] Given the heterogeneity of the field, it is also difficult to have a
analyzes the carbon impact of training their own state-of-the-art broad view of all the Green AI literature that has been published in
models. Results lead to the conclusion that we need to reduce the the past years. To understand the existing research, we conduct a
carbon footprint of developing and running AI models. systematic literature review on Green AI. We provide an overview
This self-reflection was an eye-opener to the AI research commu- and characterization of the existing research in this field. Moreover,
nity. Many papers followed, calling for a new research direction that we study how the field has been evolving over the years, pinpoint
would consider this problem. Schwartz et al. coined the term Green the main topics, approaches, artifacts, and so on.
AI as “AI research that yields novel results while taking into account This literature review shows that there has been a significant
the computational cost” [92]. Bender et al. published a position paper growth in Green AI publications – 76% of the papers have been pub-
highlighting the consequences of continuously increasing the size lished since 2020. The most popular topics revolve around monitor-
of AI models [31]. A natural question that is posed is whether we ing, hyperparameter tuning, deployment, and model benchmarking.
are doing enough as a research community to mitigate the carbon We also highlight other emerging topics that might lead to inter-
impact of developing and running AI-based software. esting solutions – namely, Data Centric Green AI, Precision/Energy
AI systems are significantly complex and, to achieve Green AI, Trade-off analysis. The current body of research has already show-
we need a joint effort that targets all the different stages of an AI cased promising results with energy savings from 13% up to 115%.
Roberto Verdecchia, June Sallou, and Luís Cruz
Still, most of the existing work focuses on the training stage of the literature indexing platforms to execute the automated search al-
AI model. Moreover, we observe that there is little involvement lows us to conduct an encompassing search of the literature based
of the industry (23%) and that most studies revolve around labora- on multiple sources, hence allowing us to mitigate potential threats
tory experiments. We argue that the field is growing to a level of to external validity, as further documented in Section 5. Following,
maturity in which involvement of the industry is quintessential to the details of each step of our research process are documented
enable the overarching goal of Green AI: harness the full potential in detail.
of AI without a negative impact in our planet. 2.2.1 Automated Initial Search. To identify a preliminary set of
To encourage open science and the reproducibility of this study, potentially relevant research works, we design an encompassing
we provide all data and scripts in a replication package available automated query to be executed on three different literature in-
online with an open-source license1 . dexers, namely Google Scholar, Scopus, and Web of Science. The
The remainder of this paper is structured as follows. In Section 2, automated query targeting publication titles states as follows:
we describe the methodology used to collect and analyze Green
AI literature. In Section 3, we present all the results yielded by our
methodology. Section 4 discusses findings and reflects on the impact
Listing 1: Automated search query
of our results in the research community. In Section 5, we reflect
on the potential threats to the validity of this study. Following, 1 INTITLE ( " green " OR " sustainab * " ) AND
Section 6 describes related work and pinpoints the differences with 2 INTITLE ( " AI " OR " ML " OR " artificial ␣ intelligence "
our study. The main conclusions and future work are presented in 3 OR " machine ␣ learning " OR " deep ␣ learning " )
Section 7.
The query is designed to retrieve literature with titles containing
2 Methodology
keywords related to sustainability, identified by the keywords green
In this section, we document the research design, which was rigor- or sustainability and its variations, e.g., “sustainable” (Listing 1,
ously adhered to during the planning and execution of the study. We Lines 1). The second part of the query instead is used to retrieve
primarily followed the guidelines for conducting SLRs in software literature concerning AI, or related synonyms and acronyms (List-
engineering research presented by Kitchenham [6]. ing 1, Lines 2-3). The query is executed on the three aforementioned
2.1 Research Objective and Question literature libraries and indexes on the 18th of July 2022, and led
to the identification of 190 potentially relevant studies. In order
The goal of this review is to understand the characteristics of ex- to be as comprehensive as possible, and avoid potential threats to
isting Green AI research. By utilizing the Goal-Question-Metric external validity, the year of publication is left unbounded in the
method [1], this objective can be described more formally as follows: automated search.
2.2.2 Application of Selection Criteria. Subsequent to the identi-
Analyze Green AI literature fication of the initial potentially relevant studies, we execute the
For the purpose of knowledge collection and categorization manual selection of the studies via a set of selection criteria defined
With respect to AI a priori. A paper is confirmed as primary study if it adheres to all
From the viewpoint of researchers and practitioners inclusion criteria, and none of the exclusion ones. The following
In the context of environmental sustainability. inclusion (I) and exclusion (E) criteria are used:
The goal of this research can be directly translated into a research
question (RQ), which states as follows: I1- The study regards AI
I2- The study regards environmental sustainability
RQ: What are the characteristics of Green AI state-of-the-art re- I3- The study regards the environmental sustainability of AI
search? I4- The study regards the software level
By answering our research question, we aim at gaining a system- E1- The study is not written in English
atic overview of the Green AI body of knowledge, starting from E2- The study is not available
an outline of the general publication trends, to a detailed analy- E3- The study is a duplicate or extensions of an already included
sis of the past and current Green AI research activities and their study
characteristics. E4- The study is a secondary or tertiary study
E5- The study is in the form of editorials, tutorials, books, ex-
2.2 Research Process tended abstracts, etc.
An overview of the research process followed is depicted in Figure 1. E6- The study is a non-scientific publication or grey literature
The process starts with the execution of a conservative automated
search query via the digital libraries and indexing platforms Google With the first three inclusion criteria (I1-I3), we ensure that the
Scholar, Scopus, and Web of Science, complemented by a subsequent primary studies focus on Green AI (I1, I2), and that the studies
iterative bidirectional snowballing process, which is conducted un- regard the environmental sustainability of AI, rather than the im-
til the achievement of theoretical saturation. Including multiple provement of environmental sustainability through AI. With the
fourth inclusion criterion instead (I4), we ensure that the primary
studies focus on software-centric Green AI. This latter criterion is
1 Replication package: https://github.com/luiscruz/slr-green-ai used to exclude studies focusing on hardware-specific Green AI
Figure 1: Systematic literature review process overview.
techniques, e.g., the use of ad hoc implemented hardware compo- The first phase consists of a data exploration process, which
nents, which we consider out of reach for most researchers/practi- terminates with the establishment of the data extraction framework
tioners interested in Green AI, and is only marginal to the definition of this study. Specifically, during this first phase, the three authors
of Green AI itself [92]. of this review independently scan the identified primary studies,
The exclusion criteria are designed to ensure that data can be and annotate the characteristics of the studies which are relevant
extracted from the papers (E1, E2), do not represent duplication or to answer our RQ. The identified characteristics are then jointly
redundancy with respect to other primary studies (E3, E4), and are discussed and refined, leading to the consolidation of the fields
provided in the form of scientific studies (E5, E6). constituting the data extraction framework of this review.
To ease the primary study selection process, adaptive reading In the second data extraction phase, the primary studies are
depth [13] is used to efficiently assess potentially relevant studies. thoroughly analyzed, and the data is extracted from the studies
In order to mitigate subjective biases and interpretations, the three according to the data extraction framework.
authors independently utilized the selection criteria to scrutinize 63- The fields of the data extraction framework utilized for this
64 candidate studies. Weekly meeting are held during the selection literature review on Green AI are the following.
process to jointly discuss examples, doubts, and align the selection
process between the three researchers. • Green AI Definition: the level of abstraction used in the paper
The application of the selection criteria concludes with the iden- to quantify the impact of AI in the surrounding environment:
tification of 16 primary studies, which constitute the starting set energy efficiency [18], carbon footprint [20], or ecological
for the subsequent snowballing process. footprint [9].
2.2.3 Snowballing. In order to enrich the set of selected primary • Study type: The overarching type of study, which could be
studies, and ensure that the primary study comprehensively repre- either presenting a position on Green AI, a Green AI solution,
sents the Green AI body of literature, the automated search results or an observational study on Green AI;
are complemented with a recursive bidirectional snowballing pro- • Topic: The Green AI topic considered in the study, e.g., hyperparameter-
cess [21]. This step entails the scrutiny of all studies either citing tuning to achieve energy efficiency of an AI algorithm;
or cited by the already included primary studies. As for the appli- • Domain: The domain considered in the study, e.g., edge or
cation of selection criteria, three researchers are involved in the mobile computing;
snowballing. During each snowballing round, the researchers in- • Type of data: The type of data utilized by AI in the study, e.g.,
dependently snowball different primary studies, and propose new text or images;
primary studies to be included, i.e., the new identified studies which • Artifact considered: The AI artifact considered in the study,
adhere to the selection criteria. During each snowballing round, ex- e.g., the data used by AI models, the AI models themselves,
amples, doubts, and divergences are jointly revisited and resolved, or the AI deployment pipeline.
and the next snowballing iteration is started. A total of two rounds • Considered phase: If the study focused on the AI training
of backward and forward snowballing are executed before no new phase, the AI inference phase, or both.
studies are identified, i.e., when theoretical saturation is reached. • Research strategy: The research strategy, as defined in [16],
The snowballing process terminates with the inclusion of 82 new used to support the claims reported in the study;
primaries studies, leading to a total of 98 primary studies which are • Dataset size: The size of the dataset, in number of data points,
considered in the literature review reported in this research. considered in the study (if any);
2.2.4 Data Extraction. In order to achieve the intended goal of • Energy Savings: The reported percentage energy savings
this study and answer our RQ (see Section 2.1), we proceed to achieved by solutions reported in the study (if any is docu-
systematically extract data from the primary studies. The data mented);
extraction process consisted of two subsequent phases. • Industry involvement: Industry involvement in the author-
ship of the study, which could be either academic-only au-
thorship, industrial-only authorship, or mixed authorship;
• Intended reader: If the study is primarily intended for aca- 33

demic readers, industrial readers, or the general public. 30
• Tool availability: The availability of the tool(s) to address
20 21
Green AI presented in the study (if any). 20
2.2.5 Data Synthesis. During the data extraction process, the data
was harmonized by relying on the constant comparison [2] of ex- 10
tracted keywords, breaking up keywords into more specific ones 6 7 7
when their semantic depth required it, or merging very similar key- 3
1
words to avoid redundancy. This analysis process relied on open 0
2015 2016 2017 2018 2019 2020 2021 2022
coding [4] to systematically identify recurrent concepts, followed
Year
by axial coding [4] to manage the increasing complexity of some
emerging concepts.
The only exceptions were made for the research strategy, in- Figure 2: Number of publications per year.
dustry involvement, and tool availability fields of the extraction
framework (see Section 2.2.4), for which provisional coding was
81
used [4]. Specifically, coding of the research strategy relied on the 75
research strategy categories reported by Stol et al. [16] was used.
The industry involvement instead relied on three pre-defined fields, 50
namely “academic-only authorship”, “industrial-only authorship”,
or “mixed authorship”. Finally, tool availability could only assume 25 20
one of two pre-defined values, namely “Yes” (if the tool is available) 9
or “No” (if the tool is not available, or none is presented in the 0
Energy Efficiency Carbon Footprint Ecological Footprint
primary study).
During the data extraction and synthesis phase, emerging codes Green AI Definition
are continuously discussed among the three authors of the review.
This process ensures that the emerging codes and their abstraction Figure 3: Number of publications per type of Green AI defi-
level are kept consistent among researchers, and are aligned with nition.
the research goal and question of the study.
3 Results
In this section, we present the results collected with our SLR on 3.3 Green AI Definition
Green AI. The distribution of publications across different Green AI definitions
is presented in Figure 3. Most literature addresses Green AI at
3.1 Publication Years the level of energy efficiency (81 papers). Higher-level definitions,
The literature spans from 2015 with the first publication on the topic namely carbon and ecological footprint, are only addressed in 20
to this present year (i.e., 2022). Figure 2 presents the distribution and 9 publications respectively. Note that a primary study might
of the literature papers regarding the publication year. We observe be mapped to more than one definition, if more than one is used in
a global increase following the years. Furthermore, a spike in the the paper at hand.
number of publications is seen in 2020, going from 7 publications
in 2019 to 20 in 2020. As the automated initial search was launched
in 2022, the publication trends reported in this review might not 3.4 Study Types
be representative of the actual research output of 2022 (see also Existing literature on Green AI spans across three types of studies,
Section 2.2.1). namely observational, solution, and position papers (see also Sec-
tion 2.2.4). As shown in Figure 4, from the 98 papers covered in
3.2 Venue Types this review, the most common are solution papers, with 51 entries,
Publications are particularly concentrated on conferences (⊲ 47 out followed by observational with 35, and position papers with 12.
of 98 papers.) and journals (⊲ 39 out of 98 papers.). Only 12 out of Note that study types are mutually exclusive, i.e., a single paper has
the 98 publications are associated with a workshop. Conferences only one study type.
being treated as an equal publishing venue as journals follows the
trends observed in the computer science research field [5, 19].
3.5 Green AI Topics
Green AI publication trends From our analysis we identify 13 main topics being addressed by the
Green AI literature. Figure 5 depicts the distribution of publications
¨ The topic of Green AI is experiencing an increasing trend
across the different topics. The most popular topic is Monitoring,
of popularity, with a considerable growth in publications from
addressed by 28 papers, followed by Hyperparameter Tuning (18),
2020 onward. Most studies are published in conferences and
Model Benchmarking (17), Deployment (17), and Model Comparison
journals, while only a minor portion in workshops.
(17). Since papers are not exclusive to a single topic, these top-4
51 Deployment ⊲ 17 out of 98 papers. Addressing the deployment

stage of the lifecycle of an AI model.
40 35 Typically, publications in this topic discuss the problem of de-
ploying AI models in a real scenario or in a scenario with pecu-
20 liar constraints that challenge a standard approach. For example,
12
deployment publications showcase the challenges of deploying
0 energy-efficient AI in FPGA [101], in Edge devices [51, 68], in mo-
Solution Observational Position bile devices [62, 79, 104], and so on.
Study type
Precision/Energy Trade Off ⊲ 11 out of 98 papers. There is a
turning point where to increase a very small fraction of the model
Figure 4: Number of publications per study type.
performance, it is required to endure an energy-intensive training
loop. Within this topic, papers address the Pareto trade-off between
topics alone cover 61% of the papers in this review. Below, we pin- having optimal accuracy and/or optimal energy efficiency.
point each topic with a short summary and the respective number Zhang et al. [118] study how removing neurons from neural
of publications. networks affects both accuracy and energy consumption. Results
indicate that a good portion of neurons are redundant and can
Monitoring ⊲ 28 out of 98 papers. Covering monitoring ap-
be removed to reduce energy consumption without a significant
proaches to study the energy and/or carbon footprint of AI models.
impact on accuracy. At the same time, it shows that there is a turning
In this topic, papers report and reflect on the energy footprint
point where removing neurons improves energy efficiency but
of state-of-the-art models throughout their lifecycle. For example,
significantly reduces accuracy. Hence, the two parameters always
Wu et al. [110] provide a landscape of the carbon footprint of AI
need to be analyzed together. Other works opt for optimizing energy
models across Facebook. Findings showcase that, typically, through-
while keeping accuracy loss within a negligible margin [106].
out the lifetime of AI models, 50% of their carbon cost lies in the
embodied carbon footprint of the hardware used to develop these Algorithm Design ⊲ 10 out of 98 papers. Design of new training
models. However, the paper shows that the vast majority of training algorithms that produce models that are significantly more energy-
workflows under-utilizes GPUs at 30–50% of their full capacity. efficient than the state of the art.
Other papers within this topic focus on solutions to make carbon Some works propose small changes to the algorithms that make a
monitoring feasible in any AI project [48]. As an example, the big difference in the final energy consumption. For example, Garcia-
Carbontracker offers a toolset to track and predict the energy and Martin et al. [46] approximate the splitting criteria by selecting
carbon footprint of training DL models [26]. These studies argue branches that require less computational effort. Results showcase
that it is quintessential to report the energy and carbon footprint decision trees that are up to 31% more energy efficient and with
of model development and training alongside performance metrics. minimal impact on accuracy. Other examples include Espnetv2 [81],
Hyperparameter Tuning ⊲ 18 out of 98 papers. Improving or a lightweight convolutional neural network designed with power-
assessing the impact on the energy consumption of optimizing efficiency in mind.
hyperparameters when training an AI model.
Many publications are motivated by the fact that tuning parame- Libraries ⊲ 8 out of 98 papers. Our choice of libraries have an
ters leads to significant energy costs – it requires retraining a model impact on the final carbon footprint of AI systems. Studies within
multiple times in order to find the optimal set of hyperparameter this topic provide some sort of evaluation of different AI libraries
values. Hence, most publications within this topic focus on iden- and how they contribute to energy efficiency.
tifying alternative strategies that reduce the number of iterations This category shows that software engineering studies play an
required to tune hyperparameters [95]. important role in enabling Green AI. Georgiou et al. [50] compare
On a different perspective, Chavannes et al. [87] explore how the energy footprint of deep learning frameworks. Results showcase
hyperparameter tuning can help deliver more energy-efficient mod- that PyTorch is more energy-efficient than Tensorflow at the train-
els by adding power consumption to the set of parameters being ing stage. However, Tensorflow tends to be more energy-efficient at
optimized. the inference stage. The study delves into the framework’s different
API methods and highlights code in the frameworks that should be
Model Benchmarking. ⊲ 17 out of 98 papers. Studies that con- optimized to reduce energy consumption. Finally, the authors moti-
tribute with benchmarks to compare the energy footprint of differ- vate the importance of reporting and discussing energy efficiency
ent models or training techniques. in the documentation of deep learning frameworks.
Benchmarks help the community understand how the state of
the art behaves w.r.t. given performance indicators. Ultimately, Data Centric ⊲ 6 out of 98 papers. Typically, the AI community
they help create baselines so that new approaches can be prop- has looked into coming up with better model training strategies.
erly validated and compared to the state of the art. As example of However, there is a new trend in AI that is raising the importance
publications within this category, Asperti et al. [27] evaluate the of developing better data collection and processing techniques as a
energy cost of different variational autoencoders. Another study, by more effective way to deliver better AI models. This line of thought
Yu et al. [117], compares the energy efficiency of common machine within Green AI aims at reducing the carbon footprint of AI by
learning algorithms when applied to clinical laboratorial datasets. tackling the problem at the data level.
Data-centric approaches for Green AI show that feature selection Green AI topics by study type
and subsampling techniques can significantly reduce the energy
consumption of training machine learning models [102]. Subsam- ¨ There are 13 main topics on Green AI. The majority (61%)
pling strategies can be more sophisticated by removing data points of the publications focuses on Monitoring, Hyperparameter-
that are expected to be redundant in terms of knowledge acquisi- tuning, Model Benchmarking, and Deployment. Despite being
tion [39]. important, topics such as Data-Centric, Estimation, and Emis-
sions are underrepresented in the scientific literature.
Network Architecture ⊲ 6 out of 98 papers. The impact of a
distributed network on the energy efficiency of AI. AI models are 3.6 Green AI Topics by Study Type
often deployed in a distributed context – e.g., IoT, edge computing, We further investigate the distribution of papers across different
etc. Hence the design and architecture of the network plays an topics per category. Figure 6 presents a bubble plot that draws a
important role in leveraging sustainable models. bubble for each pair topic (x-axis) and study type (y-axis). The size
For example, Kim and Wu [68] propose an adaptive execution of the bubble is proportional to the number of papers published in
engine that selects the inference strategy according to the signal each pair. The plot enables a few observations.
strength of the network in different devices, as it is known to affect Most topics adhere to the general pattern observed earlier in Sec-
the energy efficiency of the edge mobile system. tion 3.4: the majority of papers consist of solution studies, followed
by observational and then position. However, the topics of Model
Estimation ⊲ 5 out of 98 papers. Collecting and making sense
Benchmarking and Libraries do not follow this pattern, being mostly
of energy or climate data is far from trivial – many different factors
covered by observational papers. This is expected as these topics
contribute to the final estimation [48]. This topic revolves around
revolve around comparing different libraries and models to provide
understanding ways of estimating the energy consumption or car-
insight on the energy efficiency of different design decisions.
bon footprint of models.
Moreover, papers from the least represented topics Ethics, Policy,
Existing solutions to estimate energy consumption for software
and Emissions tend to be position papers. From the ten studies in
fail to provide meaningful insight about energy consumption that
these three topics, only one is observational and none is solution.
can be mapped to a machine learning model’s structure. IrEne cre-
Also worth noticing is the fact that the majority of the position
ates a graph that breaks down NLP models into low-level machine
studies in Green AI only cover the smallest topics. Considering
learning primitives and provides energy estimations at the primitive
the top-10 topics – from Monitoring to Estimation – only 6 are
level [37].
position papers. In contrast, the bottom-4 topics (including Other)
Emissions ⊲ 4 out of 98 papers. Papers that focus on understand- are covered by 10 position papers.
ing the carbon impact of creating and/or consuming AI systems.
Green AI topics by study type
Dhar [40] flags the importance of being able to quantify carbon im-
pact and the lack of tools and data available. Fraga-Lamas et al. [44] ¨ Most publications on Ethics, Policy, and Emissions are posi-
go beyond reporting the energy consumption of an AI-enabled tion studies calling for more research in these topics.
IoT scenario and present how much carbon would be emitted in
different countries and different energy sources. 3.7 Domains
Figure 7 presents the distribution of the publications according to
Policy ⊲ 3 out of 98 papers. Studies within this topic address and
the domain they cover. The majority of the publications (i.e., ⊲ 58
discuss strategies on how we should handle the carbon footprint of
out of 98 papers.) do not devote their studies to a specific domain, but
AI as a society.
tackle the energy efficiency of AI in a general context. Regarding
Perucica and Andjelkovic [85] reflect on the environmental poli-
the most specific studies, the most covered domains are:
cies implemented by the European Union, discussing whether they
fit the AI era or new regulations are needed. Rhode et al. [89] call Edge Regarding Internet of Things and Edge Computing, which
out for the unclear dilemma between the impact of existing/up- are usually associated with distributed systems and networks.
coming AI technologies and the commitment to achieve the 1.5℃ ⊲ 24 out of 98 papers.
climate change goal as expressed in the UNFCCC Paris Declaration. Computer Vision Regarding image recognition.⊲ 6 out of 98 pa-
pers.
Ethics ⊲ 3 out of 98 papers. Papers that focus on the ethical im- Cloud ⊲ 5 out of 98 papers.
plications of the growing carbon footprint of AI. Tamburrini [100] Mobile ⊲ 4 out of 98 papers.
discusses the responsibilities of AI scientists, AI infrastructure
providers, and other stakeholders in enabling Green AI. The pa- The Other category gathers publications about a specific domain,
per questions whether it is ethically justified to create massive AI being covered only once, among Health, Autonomous Driving,
pipelines to improve accuracy. Smart cities, Human Activity, Wearables, and Embedded Systems.
Other ⊲ 5 out of 98 papers. Studies addressing a relevant topic Green AI domains

with only a single publication in total: User values [69], Schedul- ¨ The majority of Green AI studies does not focus a specific
ing [120], Rebound Effects [109], Security [93], Energy Capping [70]. domain. Among specific domains, edge computing results to be
the most recurrent one.
28
20 18 17 17
11 10
10 8
6 6 5 5
4 3 3
0
ori
ng nin
g ing ent Off sig
n rie
s ric ure ati
on on
s
licy ics er
it -Tu ark oym ade- - De i bra Cent i t ect i m i ssi Po Eth Oth
on ter m p l r
thm
L ta -
rch Es t m
M e nch De yT Da
E
a r am e l Be n erg l g ori o r kA
per
p
Mo
d n-E A
Ne
tw
Hy e c isio
Pr
Topic
Figure 5: Number of papers per Green AI topic.
60 58
Other 2 2 1
40
Ethics 3
24
Policy 3 20
5 4 4 7
Emissions 1 3 0
General Edge Computer Cloud Mobile Other
Vision
Estimation 1 4
Domain
Network Architecture 1 5
Figure 7: Number of publications per study domain.
Data-Centric 1 4 1
Libraries 7 1 is used to make predictions from new data. Thus, we classify the
papers according to 3 categories: training, inference, and all. The
Algorithm-Design 1 9 all category translates the fact that the paper does not consider a
particular phase, but the whole pipeline.
Precision-Energy Trade-Off 3 8 As depicted in Figure 8, we find that most of the publications
on the topics of Green AI focus on the training phase (⊲ 49 out of
Deployment 5 12 98 papers.). In comparison, fewer papers direct their studies at the
inference phase (⊲ 17 out of 98 papers.) or on the overall process
(⊲ 32 out of 98 papers.).
Model Benchmarking 14 3
Hyperparameter-Tuning 6 12 49
40
Monitoring 11 12 5 32
20 17
Observational Solution Position
0
Figure 6: Number of publications by topic and study type. Training All Inference
Considered stage
3.8 AI Pipeline Phases
Figure 8: Number of publications per studied phase of AI.
The AI pipeline is divided into two major phases: the training,
when the AI model is built, and the inference, when the model
42
Green AI Pipeline Phase 40
¨ Approximately half of Green AI studies focus on the training 32
30
phase, while a minor portion considers the entire AI pipeline.
22
Only a minor portion of the Green AI literature focuses on the 20
inference phase.
10 10
4 2
3.9 Considered Artifacts
0
AI systems are based on several artifacts, and tackling the energy ef- Image Textual Numeric Video Audio Not
Specified
ficiency of such systems can thus involve multiple of those artifacts
(e.g., data, model, pipeline) or different related artifacts (e.g., archi- Data Type
tecture, framework, CPU). A distribution of the artifacts considered
in the primary studies is documented in Figure 9. The categories of Figure 10: Occurrence of data types used in the Green AI lit-
artifacts are: erature.
Model The publications within this category focus on the model
and/or associated algorithm to tackle the energy efficiency
of AI. ⊲ 63 out of 98 papers. Green AI algorithm types
Data Papers that address energy efficiency through the study of ¨ Most Green AI primary studies are algorithm-agnostic or
the data used in the AI pipeline. ⊲ 8 out of 98 papers. focus on neural networks. A small fraction uses decision trees.
Pipeline Studies looking at the whole AI pipeline. ⊲ 3 out of 98
papers. 3.11 Data Types Used
Other Publications dealing with CPU, architecture, and framework.
Regarding the types of data used in the Green AI body of literature,
⊲ 4 out of 98 papers.
an overview of their distribution is reported in Figure 10. From
General The papers do not specify a particular artifact and address
the figure, we can observe that the recurrence of data types across
AI systems as a whole. ⊲ 24 out of 98 papers.
primary studies is:
Image data ⊲ 42 out of 98 papers.
63 Textual data ⊲ 22 out of 98 papers.
60 Numeric data ⊲ 10 out of 98 papers.
Video data ⊲ 4 out of 98 papers.
Audio data ⊲ 2 out of 98 papers.
40
From the distribution of data types, we notice that image data
24 is by far the most used one, and is utilized by almost half of the
20 studies in the body of literature. The second most utilized data type
8 is textual data, which nevertheless appears approximately half as
3 4
0 often as the image one. Other types of data result to be less recurrent,
Algorithm General Data Pipeline Other with only few studies utilizing audio data (e.g., Lenherr et al. present
Artifact a metric to measure the sustainability of Green AI by considering
as case study the Intel MovidiusX processor, an embedded video
Figure 9: Number of publications per studied artifact. processor with a Neural Engine for video processing and object
detection [74]).
A rather high number of primary studies does not specify any
3.10 Algorithm Types kind of data (Not specified category, ⊲ 32 out of 98 papers.). This
finding has to be primarily attributed to the position and theoretical
By considering the primary studies which focus on a specific algo- papers included in the review (see also Section 3.4 and Section 3.5).
rithm (⊲ 51 out of 98 papers.), we note that the vast majority focus
on neural networks (⊲ 41 out of 98 papers.). Only a much smaller Green AI data types
fraction focuses on algorithms of different nature, such as decision ¨ Image data is the most used data type in Green AI studies,
trees (⊲ 5 out of 98 papers.), genetic algorithms (⊲ 1 out of 98 papers.), followed by textual and numeric data.
or logistic regression models (⊲ 5 out of 98 papers.).
Regarding the deep neural network algorithms, we also note a
further characterization of this field, with 8 studies focusing on 3.12 Dataset sizes
convolutional neural networks, one on transformers, and one Regarding the size of the datasets used in the papers, approximately
on spiking neural networks. We also observe three algorithms half of the primary studies (⊲ 48 out of 98 papers.)) directly reference
that appear only once in the Green AI literature (Other category, the number of data points used. By inspecting such numbers, we
⊲ 3 out of 98 papers.), namely genetic algorithms, logic regression note that the number of data points used to study and to evaluate
algorithms, and stochastic gradient descent algorithms. Green AI algorithms and approaches varies greatly, and ranges from
73
60
40
20
12 Academic
6 5 2
75 Industrial
20 3
0
Laboratory Field Computer Simulation Judgement None
Experiment Experiment Study
Research Strategy
Figure 11: Occurrence of research strategies used in the

Green AI literature.
1k data points [51] to 40M data points [45]. Almost half of the Figure 12: Industry involvement.
studies reporting the number of data points (⊲ 25 out of 48 papers)
utilize data points in the order of thousands (1𝑘 ≤ #𝑑𝑎𝑡𝑎𝑝𝑜𝑖𝑛𝑡𝑠 ≤
70𝑘), while the remaining (⊲ 23 out of 48 papers) use one million to optimize energy the most are based on quantizing the inputs of
data points or more (1𝑀 ≤ #𝑑𝑎𝑡𝑎𝑝𝑜𝑖𝑛𝑡𝑠 ≤ 40𝑀). decision trees [23] (97% energy savings), using data-centric Green
AI techniques [102] (92% energy savings), and leveraging efficient
Green AI dataset sizes
deployment of AI algorithms via virtualized cloud fog networks
¨ Dataset sizes range from 1k to 40M data points, with ap- (91% energy savings) [116]. Overall, more than half of the papers
proximately half of the studies utilizing 1M or more data points. explicitly reporting energy saving percentages report a saving of at
least 50% (⊲ 17 out of 27 papers), while only a minor number savings
between 13% and 49%.
3.13 Research Strategies
¨ Green AI energy savings
By considering the research strategies [16] utilized in the Green AI
literature, the distribution of the various strategies, according to Studies report energy savings between 13% and 115% energy
the collected primary studies, is reported in Figure 11. savings, with more than half of the papers reporting savings of
The majority of paper results adopt laboratory experiments at least 50%.
(⊲ 73 out of 98 papers.), while only a fraction uses other research
strategies, such as field experiments (⊲ 6 out of 98 papers.), i.e., 3.15 Industry involvement
experiments conducted in pre-existing settings and computer sim- Regarding the industry involvement in Green AI scientific publica-
ulations, i.e., “in silico” simulations conducted in a nonempirical tions (see also Section 2.2.4), an overview of the authorship of the
setting (⊲ 5 out of 98 papers.). As examples, Liu et al. [77] use a Green AI primary papers is depicted in Figure 12.
field study to assess a green software stack for computer vision From the figure, we can note that most Green AI studies are au-
of autonomous robots, while Yosuf et al. [116] leverage computer thored exclusively by academic researchers (⊲ 75 out of 98 papers.),
simulations to study how virtualized cloud fog networks can be while also a considerable portion, amounting almost to a fourth of
used to improve AI energy efficiency. The 12 papers not displaying all primary studies, are authored by a mix of academic and indus-
any research strategy correspond to the position papers (cf. the trial researchers (⊲ 20 out of 98 papers.). Green AI studies written
“None” category in Figure 11). exclusively by industrial authors appear only in rare instances (⊲ 3
out of 98 papers.).
Green AI Research Strategies
¨ Most Green AI studies use laboratory experiments, while ¨ Industry involvement
only a minority adopt other research strategies, such as field Most studies are written by academic authors, while a minor
experiments and computer simulations. portion by a mix of academic and industrial authors. Green AI
studies written exclusively by academic authors are very rare.
3.14 Energy savings
By considering the energy savings reported achievable via Green AI 3.16 Intended readers
strategies, we note that only approximately a third of the primary By considering the intended readers of the Green AI scientific
studies explicitly document them (⊲ 27 out of 98 papers.). Out of literature, the vast majority targets academic readers (⊲ 85 out
all Green AI strategies, among the ones which report concrete of 98 papers.), while a much smaller portion both academic and
saving percentages, a technique based on structure simplification industrial readers (⊲ 8 out of 98 papers.). Despite scientific papers
for deep neural networks results to save more energy, amounting targetting intuitively a specialized audience, among the Green AI
to 115% energy savings [118]. The other techniques which result literature, few studies are intended also for the general public (⊲ 5
out of 98 papers.). For example, Dhar et al. [40], present an intuitive emissions alone. Based on these considerations, we define the field
yet thoroughly positioned article on the systemic effect of AI on of Green AI as follows:
carbon emissions. Interestingly, among the primary studies, few are
intended also for policymakers, i.e., aim to sensibilize government “Green AI regards practices aimed at utilizing AI to
stakeholders to consider issues related to Green AI. For example, mitigate the impact that humans have on the natu-
in a paper by Rohde et al. [89], how opportunities and risks for ral environment in terms of natural resources utilized,
the environment, economy and society associated with AI can be and/or mitigating the impact that AI itself can have on
governed are discussed. the natural environment.”
¨ Intended readers On one hand, the definition above perfectly fits the studies fo-
The vast majority of Green AI studies are targetting academic cusing on the holistic impact that Green AI has on the natural
readers, while a much smaller portion targets both academic environment. On the other hand, given its encompassing nature,
and industrial readers. A handful of studies, especially position the definition is also suited for studies focusing on lower abstrac-
papers, are intended for the general public. tion levels of sustainability, such as Green AI 𝐶𝑂 2 emissions and
energy consumption. In the latter case however, the definition also
3.17 Tool Provision acts as a word of warning: while studying the lower levels of Green
AI is paramount, only by considering the totality of the heteroge-
Among the primary studies collected for this literature review on
neous natural resources utilized by AI can we really understand
Green AI, only a small fraction (⊲ 15 out of 98 papers.) makes tools
the environmental impact of AI.
available to tackle Green AI. The tools provided are of heteroge-
The transdisciplinary topics of Green AI (with gaps). The 13
neous nature, and range from tools to monitor the resource effi-
different topics we discover in this review emphasize that Green AI
ciency of AI algorithms [52], to tools optimizing the energy effi-
is a broad field that needs to be tackled as a transdisciplinary field.
ciency for stochastic edge inference [68], and implementations of
Some topics are naturally tied to training strategies (e.g., monitoring,
convolutional neural networks optimized for energy efficiency [81].
hyperparameter tuning, algorithm design). However, there are other
¨ Green AI Tool Provision topics that take Green AI outside the training realm.
This is the case for example of Deployment, Libraries, and Esti-
Albeit numerous studies provide solution to tackle Green AI,
mation that promise to be relevant in enabling Green AI. We argue
only a fraction of them makes tools based on the solutions
that other disciplines need to be involved. For example, Software
readily available online as an implemented tool.
Engineering which has been dealing with these topics for tradi-
tional software systems. As highlighted by Cao et al. in their work
4 Discussion on estimation [37], one cannot expect existing strategies for tradi-
The consolidated and still growing Green AI publication trend. tional software to address the new challenges of AI-based systems.
From the analysis of the publication trends a clear picture emerges. Conversely, only a few Green AI papers [50, 56, 80] come from
The topic is gaining increasing traction in the academic community, software engineering venues.
especially if the latest years are considered (from 2020 onward). Our analysis also shows that the topics Estimation and Emis-
Despite being a quite new research topic (with the first paper on sions are under-represented, with six and five papers, respectively.
Green AI being published in 2015), the socio-environmental rele- We argue that more work is quintessential in these topics to help
vance of the topic seems to be reflected in its targeted publication scientists and practitioners report the carbon footprint of their AI
venues. With conferences and journal being the most recurrent models in a seamless way.
Green AI publication venues, the Green AI research field seems We showcase that papers under the topic Policy are only covered
to have positioned and consolidated itself quite quickly within AI by position papers. We find this finding disconcerting: new policies
research communities. to encourage Green AI within both industry and academic contexts
A definition of Green AI. From the results regarding how the need to be backed up with reliable evidence. Hence, we need more
term “Green AI” is used in the literature a clear picture emerges. observational and solution papers that tackle this topic in the near
Most Green AI studies consider Green AI as exclusively related to future.
energy efficiency. Only fewer studies examine the influence of AI The same issue is present in Emissions – only one paper is
on greenhouse gas emissions (𝐶𝑂 2 ), and an even minor fraction ex- observational and the remaining are position. It might be the case
amines the holistic impact that AI has on the natural environment. that computing the climate impact of AI is far from trivial and it
By considering the different levels of abstraction (namely energy is easier said then done. Again, this is a call for the community to
efficiency, carbon footprint, and environmental footprint) the higher, take action. It is not enough to ask big companies to provide their
more encompassing level, of environmental footprint seems best data on carbon impact – we also need to provide strategies and
fitted to define the field of Green AI. In fact, as demonstrated in solutions to make it standard and straightforward.
recent literature, reducing the environmental impact of AI exclu- The fundamental Green AI research unbounded from ap-
sively to energy consumption has to be deemed as overly simplis- plication domains. From the collected results we deduce that,
tic process [8]. Similarly, as green resources are sustainable but in order to improve the environmental sustainability of AI, it is
not infinite [17], the field of Green AI has to account also for the often not necessary to focus on a specific domain. This implies
multifaceted environmental impact AI can have, other than 𝐶𝑂 2 that frequently fundamental aspects of Green AI are still open to
investigation, and results can then be ported from a generic set- of industry towards Green AI concerns, and/or the importance of
ting to specific domains. However, from the obtained results, we moving towards more environmentally sustainable AI practices.
also note that the increasing distribution of digital infrastructures As a potential impediment to the industrial adoption of Green AI
to achieve environmental sustainability [17] might have played a research, our results point to a low recurrence of studies targeted
role in Green AI research, with edge computing being the most towards practitioners. While numerous journals are explicitly aimed
considered specific domain. at practitioners, e.g., IEEE Software2 , only few studies on Green AI
The high emphasis on the AI training phase. The results included in our review target them. This result might point to the
regarding the AI pipeline phases considered in the literature un- fact that the Green AI interest is still primarily focused towards
equivocally point to training as the most studied phase. Albeit the academic activities, while the authorship showcases a rather high
training phase is intuitively the most energy-greedy phase, this re- interest of industry. As take away, similar to the considerations
sults calls for a word of caution. From recent results (e.g., a study on made for the Green AI research strategies, it might be the right
data-centric Green AI [102]) the inference phase results to consume moment to consider a higher involvement of industry in Green AI,
only a negligible fraction of the energy consumed in the training which results to date to be a research area still targeted primarily
phase. Nevertheless, given the high execution rate of the inference towards academic readers.
phase, how the energy consumed by the infrequent execution of Green AI lacks tool support. Finally, from this review, we note
the training phase compares to one of the highly executed inference that the current situation regarding the provisioning of Green AI
phase is still an open question. As a call for action, studies should tools is not bright. Albeit the majority of the studies present Green
be conducted by considering the energy consumed throughout the AI solutions, only a small fraction of them makes the solutions
whole life cycle of AI models, from their training to inference phase, available as a tool. We conjecture that this result might either point
till their eventual deprecation. towards (i) a fast-paced nature of Green AI research, in which
Image datasets as primary Green AI data source. By consid- results are rapidly deprecated, and hence tools are not meaningful,
ering the data types used in Green AI studies, we note that the or (ii) an immaturity of the research field, which still requires a
vast majority of the literature uses image data. To the best of our solid empirical foundation on which tools can be built upon.
knowledge, this choice is not guided by any specific research design
choice (e.g., AI models based on image data being the most used in 5 Threats to Validity
practice, or being the most energy greedy ones). For this reason, In this section, we discuss the threats to validity of our study. To en-
we conjecture that the popularity of utilizing image data for Green sure the quality of the results, we established a well-defined research
AI data is mostly driven by convenience, either because past work protocol to proceed with the data collection. In addition, through-
focused on such data by chance, image datasets are more accessi- out our study, we followed the recommendations of the guidelines
ble/standardized with respect to other ones, or more off the shelf for conducting a systematic literature review [6, 7, 10, 14, 21]. We
image AI models/libraries are currently available. Regardless of the designed and carried the different reviewing processes according
cause, this result points to the need of utilizing more heterogeneous to the rigorous protocol we established after the guidelines and
data types, rather than focusing primarily on image data, in order described in Section 2. Nevertheless, some threats to validity can
to gain a holistic understanding of Green AI. still exist even with our best efforts. In the following, we present
Laboratory experiments guided till now Green AI. The most the threats which could have influenced our study, jointly with the
common research strategy adopted for Green AI studies clearly strategies we adopted to mitigate them.
emerges from the literature as being laboratory experiments. Given External validity The main threat to external validity is that
the fast popularization and consolidation of the Green AI research the literature collected and analysed in this study is not sufficiently
field, from this review it seems as if the time is suitable to shift representative. To avoid this situation, we surveyed three promi-
the focus to other research strategies, e.g., field experiments and nent literature indexers through an automatic query (i.e., Google
case studies. This would not only allow to change the considered Scholar, Scopus, and Web of Science), and left the year of publica-
context from an in vitro to an in vivo setting, but also to bridge tion unbounded, to reduce the probability of missing any relevant
potential gaps between academic research and industrial practice. publication. In addition, the search query was designed to target
The highly promising energy savings of Green AI. From the relevant literature directly with specific keywords, while allow for
results of this review, we deduce that the research field of Green AI flexibility by considering similar, complementary, and variation of
is highly promising, with more than half of the papers reporting 50% the keywords (e.g., the keywords green, sustainability, and sustain-
or more energy savings. This study focuses on the state of the art able). We also mitigated the threat of having an incomplete set of
of Green AI, rather than focusing on the state of practice. It would studies, as well as the threat associated with the specificity of the
be therefore interesting to understand, as future work, the extent to terms used in the search query, by performing a complementary
which this encouraging results are transposed to industrial practice, iterative bidirectional snowballing process of the query results. This
and the potential impediments which hinder their adoption or full latter search strategy allowed us to include literature related to our
potential. query that was not directly referencing any of the automated search
A noticeable industry involvement. Regarding industry in- keywords. We limited our review of the literature to peer-reviewed
volvement in Green AI studies, the results gathered from this re- studies, to moderate the threat about the low quality of the set of
view are promising. The authorship of Green AI literature results primary studies. We deem that such practice does not constitute
to be to a good extent shared between academic and industrial re-
searchers/practitioners. This finding might highlight the sensibility 2 https://www.computer.org/csdl/magazine/so. Accessed 22nd December 2022.
an additional threat, as peer-review is a standard requirement of primary studies (98 versus 41 papers). This difference could be ex-
high-quality publications. plained by the fact that their review only includes papers involving
Internal validity To address potential threats to internal valid- consumer products and services and excludes papers dealing with
ity, we established a rigorous research protocol a priori, and we non-commercial applications, whereas we provide an overview of
followed it to conduct all the research activities. Subjective biases the whole field of Green AI.
and interpretations were mitigated by closely complying with the Previous literature reviews consider Green AI research by fo-
selection criteria to evaluate the studies. Moreover, weekly meeting cusing exclusively on specific subdomains of AI and application
were held during the selection process to jointly discuss exam- subdomains of Software Engineering, e.g., deep learning [22], in-
ples, doubts, and to align the selection process between the three formation retrieval [15], or embedded systems [11]. In contrast,
researchers. our research aims to review the entirety of the Green AI literature,
Construct validity To ensure that the set of studies answered regardless of the specific AI or software engineering subdomain it
our research questions, we applied a priori carefully constructed focuses on.
inclusion and exclusion criteria to strictly control the manual selec- In the survey of Xu et al. [22], the authors provide an overview
tion of studies. We then used the bidirectional snowballing tech- of the approaches aimed at improving the environmental sustain-
nique to expand the range of relevant primary studies to a more ability of deep learning. The authors map the different approaches
comprehensive set. using a taxonomy of the deep learning life cycle stage and its related
Conclusion validity Possible sources of bias arising from the artifacts. In contrast to such study, in this review we target a higher
data extraction and analysis phases were mitigated by strict compli- number of Green AI characteristics (see Section 2.2.4), and target
ance with an a priori defined protocol, explicitly tailored to collect the entirety of Green AI literature, rather than exclusively the one
the data needed to answer our research questions. In all, we fol- on deep learning.
lowed the best practises of the standard guidelines for systematic Scells et al. [15] provide a literature review on methods related
literature reviews [6, 7, 10, 14, 21]. Lastly, we documented all the to the domain of Green Information Retrieval. The authors explain
data throughout the whole review process and made them available that the domain of Information Retrieval (IR) produces relatively
for reproducibility and replicability purposes (see Section 1). low emissions compared to other research domains, but they also
warn that similar trends of costs and environmental impact may
6 Related Work appear considering the growing development of new IR-focused
Despite the growing interest around Green AI, the topic has been deep learning models. Natural Language Processing and Machine
marginally considered only in a handful of reviews. The related Learning are also discussed, but only with respect to the Information
work manly investigates the topic as an intersection of AI and envi- Retrieval domain. Therefore, they are not addressing the whole field
ronmental sustainability, or by defining it as a specific subdomain of AI, as done in this review.
of software engineering. To the best of our knowledge, this review Finally, the optimizations that can be made for the implementa-
is the first aiming towards a comprehensive review of Green AI tion of deep learning models on the specific platform of NVIDIA
research and its characteristics. Jetson are reviewed with a focus on energy efficiency by Mittal [11].
In a recent publication, Natarajan et al. perform a systematic The review covers studies at both the hardware and software level.
literature review on the topics of ‘AI for Environmental Sustain- Nevertheless, the review addresses only the Jetson platform 3 . We
ability’ as well as ‘Environmental Sustainability of AI’. The authors differentiate ourselves from this study by providing a holistic review
present the affordances of the use of AI for sustainability that they of Green AI, rather than focusing exclusively on deep learning.
extracted from the literature [12]. ‘AI affordances’ are introduced
as the posible actions offered by AI artifacts to an organizational 7 Conclusion
actor whose goal is to achieve environmental sustainability. The In this systematic literature review, we aimed at characterizing
authors point out the focus of previous research on the technical the existing body of research in Green AI. We identified 98 peer-
side, and they advocate for a further exploration of the concept reviewed publications that show a significant growth in this re-
of sustainable AI affordances from a socio-technical perspective. search field since 2020.
The literature is exclusively analyzed with respect to building the We provide an encompassing overview and characterization of
AI affordances, and other characteristics of the state-of-the art of the different topics being addressed by Green AI papers. We identi-
Green AI are considered nor discussed in the study. In contrast, fied 13 different Green AI topics, showcasing that the spotlight falls
our review focuses on the sustainability of AI, and maps the en- on monitoring, hyperparameter-tuning, model benchmarking, and
tirety of the Green AI literature. In our review, we aim at providing deployment. Less frequent topics – such as data-centric, estimation,
a detailed and comprehensive overview of the characteristics of and emissions – show less obvious approaches that deserve further
the Green AI state-of-the-art research (e.g., topic, domain, type research in the upcoming years.
of study, targeted artifact, overview of energy savings, tool provi- The potential of Green AI cannot be disregarded: the majority of
sion, industrial involvement). Therefore, in contrast to the work publications show significant energy savings, up to 115%, at little
of Natarajan et al. [12], we consider the different facets of Green or no cost in accuracy. However, we argue that most publications
AI, rather than exclusively on AI affordances, leading to a more revolve around laboratory studies. More field experiments are quin-
holistic review of Green AI, and a higher number of considered tessential to help AI practitioners embrace green strategies that
3 https://developer.nvidia.com/embedded-computing. Accessed 23th December 2022.

are effective, feasible, and mensurable. This is also reflected in the //doi.org/10.1002/asi.23349
small participation of the industry in these studies – only 23% of [20] Thomas Wiedmann and Jan Minx. 2008. A definition of ‘carbon footprint’.
Ecological economics research trends 1, 2008 (2008), 1–11.
publications involve industry partners. [21] Claes Wohlin. 2014. Guidelines for snowballing in systematic literature stud-
At the same time, we conclude that the field seems to be reaching ies and a replication in software engineering. In International Conference on
Evaluation and Assessment in Software Engineering. ACM Press, 1–10.
a considerable level of maturity. Hence, it is necessary to encourage [22] Jingjing Xu, Wangchunshu Zhou, Zhiyi Fu, Hao Zhou, and Lei Li. 2021. A Survey
the port of promising academic results to industrial practice. In on Green Deep Learning. arXiv (Nov. 2021). https://doi.org/10.48550/arXiv.2111.
other words, our study calls out for the importance of having re- 05193 arXiv:2111.05193
producible research. Only a small fraction of solution papers offers
a tool or software package that can be used by the community. We
Primary Studies
argue that Green AI is an urgent and necessary line of research [23] Brunno Abreu, Mateus Grellert, and Sergio Bampi. 2020. VLSI design of tree-
based inference for low-power learning applications. In 2020 IEEE International
that needs to grow fast and solid – non-replicable research can only Symposium on Circuits and Systems (ISCAS). IEEE, 1–5.
slow us down. [24] Brunno Abreu, Mateus Grellert, and Sergio Bampi. 2022. A framework for design-
This review also serves as a foundation for future research that ing power-efficient inference accelerators in tree-based learning applications.
Engineering Applications of Artificial Intelligence 109 (2022), 104638.
ultimately aims to reduce the climate impact of AI. In this respect, [25] Phyllis Ang, Bhuwan Dhingra, and Lisa Wu Wills. 2022. Characterizing the
we see potential in follow-up grey literature or interview studies Efficiency vs. Accuracy Trade-off for Long-Context NLP Models. In Proceedings
of NLP Power! The First Workshop on Efficient Benchmarking in NLP. Association
to understand how AI professionals are currently addressing the for Computational Linguistics, Dublin, Ireland, 113–121. https://doi.org/10.
issue. 18653/v1/2022.nlppower-1.12
[26] Lasse F. Wolff Anthony, Benjamin Kanding, and Raghavendra Selvan. 2020.
References Carbontracker: Tracking and Predicting the Carbon Footprint of Training Deep
Learning Models. ICML Workshop on Challenges in Deploying and monitoring
[1] Victor R. Basili, Gianluigi Caldiera, and Dieter Rombach. 1994. The Goal Question Machine Learning Systems. arXiv:2007.03051.
Metric Approach. In Encyclopedia of Software Engineering. Wiley, 528–532. [27] Andrea Asperti, Davide Evangelista, and Elena Loli Piccolomini. 2021. A Survey
[2] Barney G Glaser. 1965. The constant comparative method of qualitative analysis. on Variational Autoencoders from a Green AI Perspective. SN Comput. Sci. 2, 4
Social problems 12, 4 (1965), 436–445. (July 2021), 1–23. https://doi.org/10.1007/s42979-021-00702-9
[3] Mark Haakman, Luís Cruz, Hennie Huijgens, and Arie van Deursen. 2021. AI [28] Nesrine Bannour, Sahar Ghannay, Aurélie Névéol, and Anne-Laure Ligozat.
lifecycle models need to be revised. Empirical Software Engineering 26, 5 (2021), 2021. Evaluating the carbon footprint of NLP methods: a survey and analysis of
1–29. existing tools. In EMNLP, Workshop SustaiNLP.
[4] Bryan Jenner, Uwe Flick, Ernst von Kardoff, and Ines Steinke. 2004. A companion [29] Michel Barlaud and Frédéric Guyard. 2021. Learning sparse deep neural net-
to qualitative research. Sage, 271–275. works using efficient structured projections on convex constraints for green AI.
[5] Jinseok Kim. 2019. Author-based analysis of conference versus journal publication In 2020 25th International Conference on Pattern Recognition (ICPR). 1566–1573.
in computer science. J. Assoc. Inf. Sci. Technol. 70, 1 (Jan. 2019), 71–82. https: https://doi.org/10.1109/ICPR48806.2021.9412162
//doi.org/10.1002/asi.24079 [30] Soroush Bateni, Husheng Zhou, Yuankun Zhu, and Cong Liu. 2018. Predjoule:
[6] Barbara Kitchenham. 2004. Procedures for performing systematic reviews. Keele, A timing-predictable energy optimization framework for deep neural networks.
UK, Keele University 33, TR/SE-0401 (2004), 28. In 2018 IEEE Real-Time Systems Symposium (RTSS). IEEE, 107–118.
[7] Barbara Kitchenham and Pearl Brereton. 2013. A systematic review of systematic [31] Emily M. Bender, Timnit Gebru, Angelina McMillan-Major, and Shmargaret
review process research in software engineering. Information and software Shmitchell. 2021. On the Dangers of Stochastic Parrots: Can Language Models
technology 55, 12 (2013), 2049–2075. Be Too Big? In FAccT ’21: Proceedings of the 2021 ACM Conference on Fairness,
[8] Alexandra Sasha Luccioni and Alex Hernandez-Garcia. 2023. Counting Carbon: A Accountability, and Transparency. Association for Computing Machinery, New
Survey of Factors Influencing the Emissions of Machine Learning. arXiv preprint York, NY, USA, 610–623. https://doi.org/10.1145/3442188.3445922
arXiv:2302.08476 (2023). [32] Alexander EI Brownlee, Jason Adair, Saemundur O Haraldsson, and John Jabbo.
[9] Jan Matuštík and Vladimír Kočí. 2021. What is a footprint? A conceptual analysis 2021. Exploring the accuracy–energy trade-off in machine learning. In 2021
of environmental footprint indicators. Journal of Cleaner Production 285 (2021), IEEE/ACM International Workshop on Genetic Improvement (GI). IEEE, 11–18.
124833. [33] Sevda Ozge Bursa, Ozlem Durmaz Incel, and Gulfem Isiklar Alptekin. 2022.
[10] Philipp Mayring et al. 2004. Qualitative content analysis. A companion to Transforming Deep Learning Models for Resource-Efficient Activity Recognition
qualitative research 1, 2 (2004), 159–176. on Mobile Devices. In 2022 5th Conference on Cloud and Internet of Things (CIoT).
[11] Sparsh Mittal. 2019. A Survey on optimized implementation of deep learning IEEE, 83–89.
models on the NVIDIA Jetson platform. J. Syst. Archit. 97 (Aug. 2019), 428–442. [34] Ermao Cai, Da-Cheng Juan, Dimitrios Stamoulis, and Diana Marculescu. 2017.
https://doi.org/10.1016/j.sysarc.2019.01.011 NeuralPower: Predict and Deploy Energy-Efficient Convolutional Neural Net-
[12] Harish Karthi Natarajan, Danielly de Paula, Christian Dremel, and Falk Ueber- works. In Proceedings of the Ninth Asian Conference on Machine Learning (Pro-
nickel. 2022. A Theoretical Review on AI Affordances for Sustainability. In ceedings of Machine Learning Research, Vol. 77), Min-Ling Zhang and Yung-
Americas Conference on Information Systems. Kyun Noh (Eds.). PMLR, Yonsei University, Seoul, Republic of Korea, 622–637.
[13] Kai Petersen, Robert Feldt, Shahid Mujtaba, and Michael Mattsson. 2008. Sys- https://proceedings.mlr.press/v77/cai17a.html
tematic mapping studies in software engineering. International Conference on [35] Antonio Candelieri, Riccardo Perego, and Francesco Archetti. 2021. Green
Evaluation and Assessment in Software Engineering, 68–77. machine learning via augmented Gaussian processes and multi-information
[14] Kai Petersen, Sairam Vakkalanka, and Ludwik Kuzniarz. 2015. Guidelines for source optimization. Soft Comput. 25, 19 (Oct. 2021), 12591–12603. https:
conducting systematic mapping studies in software engineering: An update. //doi.org/10.1007/s00500-021-05684-7
Information and software technology 64 (2015), 1–18. [36] Qingqing Cao, Aruna Balasubramanian, and Niranjan Balasubramanian. 2020.
[15] Harrisen Scells, Shengyao Zhuang, and Guido Zuccon. 2022. Reduce, Reuse, Towards Accurate and Reliable Energy Measurement of NLP Models. In Pro-
Recycle: Green Information Retrieval Research. In SIGIR ’22: Proceedings of ceedings of SustaiNLP: Workshop on Simple and Efficient Natural Language Pro-
the 45th International ACM SIGIR Conference on Research and Development in cessing. Association for Computational Linguistics, Online, 141–148. https:
Information Retrieval. Association for Computing Machinery, New York, NY, USA, //doi.org/10.18653/v1/2020.sustainlp-1.19
2825–2837. https://doi.org/10.1145/3477495.3531766 [37] Qingqing Cao, Yash Kumar Lal, Harsh Trivedi, Aruna Balasubramanian, and
[16] Klaas-Jan Stol and Brian Fitzgerald. 2018. The ABC of software engineering Niranjan Balasubramanian. 2021. IrEne: Interpretable Energy Prediction for
research. ACM Transactions on Software Engineering and Methodology (TOSEM) Transformers. Proceedings of the 59th Annual Meeting of the Association for
27, 3 (2018), 1–51. Computational Linguistics and the 11th International Joint Conference on Natural
[17] Roberto Verdecchia, Patricia Lago, and Carol de Vries. 2022. The future of Language Processing (Volume 1: Long Papers) (2021). https://doi.org/10.18653/
sustainable digital infrastructures: A landscape of solutions, adoption factors, v1/2021.acl-long.167
impediments, open problems, and scenarios. Sustainable Computing: Informatics [38] Francisco M Castro, Nicolás Guil, Manuel J Marín-Jiménez, Jesús Pérez-Serrano,
and Systems (2022), 100767. and Manuel Ujaldón. 2019. Energy-based tuning of convolutional neural net-
[18] Roberto Verdecchia, Patricia Lago, Christof Ebert, and Carol De Vries. 2021. Green works on multi-GPUs. Concurrency and Computation: Practice and Experience
IT and Green Software. IEEE Software 38, 6 (2021), 7–15. 31, 21 (2019), e4786.
[19] George Vrettas and Mark Sanderson. 2015. Conferences versus journals in
computer science. J. Assoc. Inf. Sci. Technol. 66, 12 (Dec. 2015), 2674–2684. https:
[39] Priyadarshan Dhabe, Param Mirani, Rahul Chugwani, and Sadanand Gande- Conference & Exhibition (DATE), 2017. IEEE, 1474–1479.
war. 2021. Data Set Reduction to Improve Computing Efficiency and Energy [59] Peter Henderson, Jieru Hu, Joshua Romoff, Emma Brunskill, Dan Jurafsky, and
Consumption in Healthcare Domain. In Digital Literacy and Socio-Cultural Joelle Pineau. 2020. Towards the Systematic Reporting of the Energy and Carbon
Acceptance of ICT in Developing Countries. Springer, 53–64. Footprints of Machine Learning. Journal of Machine Learning Research 21, 248
[40] Payal Dhar. 2020. The carbon impact of artificial intelligence. Nat. Mach. Intell. (2020), 1–43. http://jmlr.org/papers/v21/20-312.html
2, 8 (2020), 423–425. [60] Miro Hodak and Ajay Dholakia. 2021. Recent Efficiency Gains in Deep Learning:
[41] Josefa Díaz-Álvarez, Pedro A Castillo, Francisco Fernández de Vega, Fran- Performance, Power, and Sustainability. In 2021 IEEE International Conference
cisco Chávez, and Jorge Alvarado. 2022. Population size influence on the en- on Big Data (Big Data). IEEE, 2040–2045.
ergy consumption of genetic programming. Measurement and Control (2022), [61] Hamzaoui Ikhlasse, Duthil Benjamin, Courboulay Vincent, and Medromi
00202940211064471. Hicham. 2022. Recent implications towards sustainable and energy efficient
[42] Jesse Dodge, Taylor Prewitt, Remi Tachet des Combes, Erika Odmark, Roy AI and big data implementations in cloud-fog systems: A newsworthy inquiry.
Schwartz, Emma Strubell, Alexandra Sasha Luccioni, Noah A. Smith, Nicole Journal of King Saud University - Computer and Information Sciences 34, 10, Part
DeCario, and Will Buchanan. 2022. Measuring the Carbon Intensity of AI in A (Nov. 2022), 8867–8887. https://doi.org/10.1016/j.jksuci.2021.11.002
Cloud Instances. In FAccT ’22: 2022 ACM Conference on Fairness, Accountability, [62] Nitthilan Kanappan Jayakodi, Syrine Belakaria, Aryan Deshwal, and Janard-
and Transparency. Association for Computing Machinery, New York, NY, USA, han Rao Doppa. 2020. Design and optimization of energy-accuracy tradeoff
1877–1894. https://doi.org/10.1145/3531146.3533234 networks for mobile platforms via pretrained deep models. ACM Transactions
[43] Mariza Ferro, Gabrieli D. Silva, Felipe B. de Paula, Vitor Vieira, and Bruno on Embedded Computing Systems (TECS) 19, 1 (2020), 1–24.
Schulze. 2021. Towards a sustainable artificial intelligence: A case study of [63] Wandri Jooste, Rejwanul Haque, and Andy Way. 2022. Knowledge Distillation:
energy efficiency in decision tree algorithms. Concurrency Computat. Pract. A Method for Making Neural Machine Translation More Efficient. Information
Exper. n/a, n/a (Dec. 2021), e6815. https://doi.org/10.1002/cpe.6815 13, 2 (Feb. 2022), 88. https://doi.org/10.3390/info13020088
[44] Paula Fraga-Lamas, Sérgio Ivan Lopes, and Tiago M. Fernández-Caramés. 2021. [64] Sorin Liviu Jurj, Flavius Opritoiu, and Mircea Vladutiu. 2020. Environmentally-
Green IoT and Edge AI as Key Technological Enablers for a Sustainable Digital friendly metrics for evaluating the performance of deep learning models and
Transition towards a Smart Circular Economy: An Industry 5.0 Use Case. Sensors systems. In International Conference on Neural Information Processing. Springer,
21, 17 (Aug. 2021), 5745. https://doi.org/10.3390/s21175745 232–244.
[45] Eva Garcia-Martin, Niklas Lavesson, and Håkan Grahn. 2017. Identification [65] Petra Jääskeläinen, Daniel Pargman, and André Holzapfel. 2022. On the envi-
of Energy Hotspots: A Case Study of the Very Fast Decision Tree. In Green, ronmental sustainability of AI art(s). In Eighth Workshop on Computing within
Pervasive, and Cloud Computing. Springer, Cham, Switzerland, 267–281. https: Limits. https://doi.org/10.21428/bf6fb269.c46375fa
//doi.org/10.1007/978-3-319-57186-7_21 [66] Lynn H. Kaack, Priya L. Donti, Emma Strubell, George Kamiya, Felix Creutzig,
[46] Eva García-Martín, Niklas Lavesson, Håkan Grahn, Emiliano Casalicchio, and and David Rolnick. 2022. Aligning artificial intelligence with climate change
Veselka Boeva. 2021. Energy-aware very fast decision tree. Int. J. Data Sci. Anal. mitigation. Nat. Clim. Change 12 (June 2022), 518–527. https://doi.org/10.1038/
11, 2 (March 2021), 105–126. https://doi.org/10.1007/s41060-021-00246-4 s41558-022-01377-7
[47] Ángel M. García-Vico and Francisco Herrera. 2021. A Preliminary Analysis [67] Minsu Kim, Walid Saad, Mohammad Mozaffari, and Merouane Debbah. 2022. On
on Software Frameworks for the Development of Spiking Neural Networks. the Tradeoff between Energy, Precision, and Accuracy in Federated Quantized
In Hybrid Artificial Intelligent Systems. Springer, Cham, Switzerland, 564–575. Neural Networks. In ICC 2022 - IEEE International Conference on Communications.
https://doi.org/10.1007/978-3-030-86271-8_47 2194–2199. https://doi.org/10.1109/ICC45855.2022.9838362
[48] Eva García-Martín, Crefeda Faviola Rodrigues, Graham Riley, and Håkan Grahn. [68] Young Geun Kim and Carole-Jean Wu. 2020. Autoscale: Energy efficiency
2019. Estimation of energy consumption in machine learning. J. Parallel and optimization for stochastic edge inference using reinforcement learning. In 2020
Distrib. Comput. 134 (2019), 75–88. 53rd Annual IEEE/ACM International Symposium on Microarchitecture (MICRO).
[49] Kent Gauen, Rohit Rangan, Anup Mohan, Yung-Hsiang Lu, Wei Liu, and IEEE, 1082–1096.
Alexander C. Berg. 2017. Low-power image recognition challenge. In 2017 [69] Pascal D. König, Stefan Wurster, and Markus B. Siewert. 2022. Consumers
22nd Asia and South Pacific Design Automation Conference (ASP-DAC). 99–104. are willing to pay a price for explainable, but not for green AI. Evidence
https://doi.org/10.1109/ASPDAC.2017.7858303 from a choice-based conjoint analysis. Big Data & Society 9, 1 (Jan. 2022),
[50] Stefanos Georgiou, Maria Kechagia, Tushar Sharma, Federica Sarro, and Ying 20539517211069632. https://doi.org/10.1177/20539517211069632
Zou. 2022. Green AI: do deep learning frameworks have different costs? In [70] Adam Krzywaniak, Pawel Czarnul, and Jerzy Proficz. 2022. GPU Power Capping
ICSE ’22: Proceedings of the 44th International Conference on Software Engineering. for Energy-Performance Trade-Offs in Training of Deep Convolutional Neural
Association for Computing Machinery, New York, NY, USA, 1082–1094. https: Networks for Image Recognition. In International Conference on Computational
//doi.org/10.1145/3510003.3510221 Science. Springer, 667–681.
[51] Santosh Gondi and Vineel Pratap. 2021. Performance and Efficiency Evaluation [71] Mohit Kumar, Xingzhou Zhang, Liangkai Liu, Yifan Wang, and Weisong Shi.
of ASR Inference on the Edge. Sustainability 13, 22 (Nov. 2021), 12392. https: 2020. Energy-Efficient Machine Learning on the Edges. In 2020 IEEE International
//doi.org/10.3390/su132212392 Parallel and Distributed Processing Symposium Workshops (IPDPSW). 912–921.
[52] Achim Guldner, Sandro Kreten, and Stefan Naumann. 2021. Exploration and https://doi.org/10.1109/IPDPSW50202.2020.00153
systematic assessment of the resource efficiency of Machine Learning.. In GI- [72] Jaeha Kung, Duckhwan Kim, and Saibal Mukhopadhyay. 2015. A power-aware
Jahrestagung. 287–299. digital feedforward neural network platform with backpropagation driven ap-
[53] Abhishek Gupta, Camylle Lanteigne, and Sara Kingsley. 2020. SECure: A So- proximate synapses. In 2015 IEEE/ACM International Symposium on Low Power
cial and Environmental Certificate for AI Systems. ICML 2020 Challenges in Electronics and Design (ISLPED). IEEE, 85–90.
Deploying and monitoring Machine Learning Systems Workshop (June 2020). [73] Loïc Lannelongue, Jason Grealey, and Michael Inouye. 2021. Green Algorithms:
https://doi.org/10.48550/arXiv.2006.06217 arXiv:2006.06217 Quantifying the Carbon Footprint of Computation. Adv. Sci. 8, 12 (June 2021),
[54] María Gutiérrez, Ma Ángeles Moraga, and Félix García. 2022. Analysing the 2100707. https://doi.org/10.1002/advs.202100707
energy impact of different optimisations for machine learning models. In 2022 [74] Nicola Lenherr, René Pawlitzek, and Bruno Michel. 2021. New universal sustain-
International Conference on ICT for Sustainability (ICT4S). IEEE, 46–52. ability metrics to assess edge intelligence. Sustainable Computing: Informatics
[55] Başak Güler and Aylin Yener. 2021. Energy-Harvesting Distributed Machine and Systems 31 (Sept. 2021), 100580. https://doi.org/10.1016/j.suscom.2021.
Learning. In 2021 IEEE International Symposium on Information Theory (ISIT). 100580
320–325. https://doi.org/10.1109/ISIT45174.2021.9518045 [75] Da Li, Xinbo Chen, Michela Becchi, and Ziliang Zong. 2016. Evaluating the
[56] Raluca Maria Hampau, Maurits Kaptein, Robin van Emden, Thomas Rost, and Energy Efficiency of Deep Convolutional Neural Networks on CPUs and GPUs.
Ivano Malavolta. 2022. An Empirical Study on the Performance and Energy In 2016 IEEE International Conferences on Big Data and Cloud Computing (BD-
Consumption of AI Containerization Strategies for Computer-Vision Tasks Cloud), Social Computing and Networking (SocialCom), Sustainable Computing
on the Edge. In Proceedings of the International Conference on Evaluation and and Communications (SustainCom) (BDCloud-SocialCom-SustainCom). 477–484.
Assessment in Software Engineering 2022 (Gothenburg, Sweden) (EASE ’22). https://doi.org/10.1109/BDCloud-SocialCom-SustainCom.2016.76
Association for Computing Machinery, New York, NY, USA, 50–59. https: [76] Anne-Laure Ligozat, Julien Lefevre, Aurélie Bugeau, and Jacques Combaz. 2022.
//doi.org/10.1145/3530019.3530025 Unraveling the Hidden Environmental Impacts of AI Solutions for Environment
[57] Walid A. Hanafy, Tergel Molom-Ochir, and Rohan Shenoy. 2021. Design Con- Life Cycle Assessment of AI Solutions. Sustainability 14, 9 (April 2022), 5172.
siderations for Energy-efficient Inference on Edge Devices. In e-Energy ’21: https://doi.org/10.3390/su14095172
Proceedings of the Twelfth ACM International Conference on Future Energy Sys- [77] Liangkai Liu, Jiamin Chen, Marco Brocanelli, and Weisong Shi. 2019. E2M: an
tems. Association for Computing Machinery, New York, NY, USA, 302–308. energy-efficient middleware for computer vision applications on autonomous
https://doi.org/10.1145/3447555.3465326 mobile robots. In Proceedings of the 4th ACM/IEEE Symposium on Edge Computing.
[58] Soheil Hashemi, Nicholas Anthony, Hokchhay Tann, R Iris Bahar, and Sherief 59–73.
Reda. 2017. Understanding the impact of precision quantization on the accu- [78] Michele Magno, Michael Pritz, Philipp Mayer, and Luca Benini. 2017. DeepEmote:
racy and energy of neural networks. In Design, Automation & Test in Europe Towards multi-layer neural networks in a low power wearable multi-sensors
bracelet. In 2017 7th IEEE International Workshop on Advances in Sensors and Networks on CPUs and GPUs. In National Conference of Theoretical Computer
Interfaces (IWASI). IEEE, 32–37. Science. Springer, 196–221.
[79] Susmita Dey Manasi, Farhana Sharmin Snigdha, and Sachin S Sapatnekar. 2020. [99] Yuxuan Sun, Sheng Zhou, and Deniz Gündüz. 2020. Energy-aware analog
NeuPart: Using analytical models to drive energy-efficient partitioning of CNN aggregation for federated learning with redundant data. In ICC 2020-2020 IEEE
computations on cloud-connected mobile clients. IEEE Transactions on Very International Conference on Communications (ICC). IEEE, 1–7.
Large Scale Integration (VLSI) Systems 28, 8 (2020), 1844–1857. [100] Guglielmo Tamburrini. 2022. The AI Carbon Footprint and Responsibilities
[80] Andrea McIntosh, Safwat Hassan, and Abram Hindle. 2019. What can Android of AI Scientists. Philosophies 7, 1 (Jan. 2022), 4. https://doi.org/10.3390/
mobile app developers do about the energy consumption of machine learning? philosophies7010004
Empir. Software Eng. 24, 2 (April 2019), 562–601. https://doi.org/10.1007/s10664- [101] Yudong Tao, Rui Ma, Mei-Ling Shyu, and Shu-Ching Chen. 2020. Challenges
018-9629-2 in Energy-Efficient Deep Neural Network Training With FPGA. In Proceedings
[81] Sachin Mehta, Mohammad Rastegari, Linda Shapiro, and Hannaneh Hajishirzi. of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)
2019. Espnetv2: A light-weight, power efficient, and general purpose convolu- Workshops.
tional neural network. In Proceedings of the IEEE/CVF conference on computer [102] Roberto Verdecchia, Luís Cruz, June Sallou, Michelle Lin, James Wickenden,
vision and pattern recognition. 9190–9200. and Estelle Hotellier. 2022. Data-Centric Green AI An Exploratory Empirical
[82] Thaha Mohammed, Aiiad Albeshri, Iyad Katib, and Rashid Mehmood. 2020. Study. In 2022 International Conference on ICT for Sustainability (ICT4S). IEEE,
UbiPriSEQ—Deep Reinforcement Learning to Manage Privacy, Security, Energy, 35–45. https://doi.org/10.1109/ICT4S55073.2022.00015
and QoS in 5G IoT HetNets. Appl. Sci. 10, 20 (Oct. 2020), 7120. https://doi.org/ [103] Chengcheng Wan, Muhammad Santriaji, Eri Rogers, Henry Hoffmann, Michael
10.3390/app10207120 Maire, and Shan Lu. 2020. {ALERT }: Accurate learning for energy and timeli-
[83] Elena Morotti, Davide Evangelista, and Elena Loli Piccolomini. 2021. A green ness. In 2020 USENIX Annual Technical Conference (USENIX ATC 20). 353–369.
prospective for learned post-processing in sparse-view tomographic reconstruc- [104] Cong Wang, Bin Hu, and Hongyi Wu. 2022. Energy Minimization for Federated
tion. Journal of Imaging 7, 8 (2021), 139. Asynchronous Learning on Battery-Powered Mobile Devices via Application
[84] David Patterson, Joseph Gonzalez, Urs Hölzle, Quoc Le, Chen Liang, Lluis- Co-running. In 2022 IEEE 42nd International Conference on Distributed Computing
Miquel Munguia, Daniel Rothchild, David R. So, Maud Texier, and Jeff Dean. Systems (ICDCS). 939–949. https://doi.org/10.1109/ICDCS54860.2022.00095
2022. The Carbon Footprint of Machine Learning Training Will Plateau, Then [105] Qu Wang, Yong Xiao, Huixiang Zhu, Zijian Sun, Yingyu Li, and Xiaohu Ge. 2021.
Shrink. Computer 55, 7 (2022), 18–28. https://doi.org/10.1109/MC.2022.3148714 Towards Energy-efficient Federated Edge Intelligence for IoT Networks. In 2021
[85] Natasa Perucica and Katarina Andjelkovic. 2022. Is the future of AI sustainable? IEEE 41st International Conference on Distributed Computing Systems Workshops
A case study of the European Union. Transforming Government: People, Process (ICDCSW). 55–62. https://doi.org/10.1109/ICDCSW53096.2021.00016
and Policy 16, 3 (June 2022), 347–358. https://doi.org/10.1108/TG-06-2021-0106 [106] Yu Wang, Rong Ge, and Shuang Qiu. 2020. Energy-Aware DNN Graph Opti-
[86] Supadchaya Puangpontip and Rattikorn Hewett. 2020. Energy Usage of Deep mization. Resource-Constrained Machine Learning (ReCoML) Workshop of ML-
Learning in Smart Cities. In 2020 International Conference on Computational Sys 2020 Conference (May 2020). https://doi.org/10.48550/arXiv.2005.05837
Science and Computational Intelligence (CSCI). IEEE, 1143–1148. arXiv:2005.05837
[87] Lucas Høyberg Puvis de Chavannes, Mads Guldborg Kjeldgaard Kongsbak, [107] Yue Wang, Ziyu Jiang, Xiaohan Chen, Pengfei Xu, Yang Zhao, Yingyan Lin,
Timmie Rantzau, and Leon Derczynski. 2021. Hyperparameter Power Impact in and Zhangyang Wang. 2019. E2-Train: Training State-of-the-art CNNs with
Transformer Language Model Training. In Proceedings of the Second Workshop on Over 80% Energy Savings. In Advances in Neural Information Processing Systems,
Simple and Efficient Natural Language Processing. Association for Computational H. Wallach, H. Larochelle, A. Beygelzimer, F. d'Alché-Buc, E. Fox, and R. Garnett
Linguistics, Virtual, 96–118. https://doi.org/10.18653/v1/2021.sustainlp-1.12 (Eds.), Vol. 32. Curran Associates, Inc. https://proceedings.neurips.cc/paper/
[88] Crefeda Faviola Rodrigues, Graham Riley, and Mikel Luján. 2018. SyNERGY: 2019/file/663772ea088360f95bac3dc7ffb841be-Paper.pdf
An energy measurement and prediction framework for Convolutional Neural [108] Simon Wenninger, Can Kaymakci, Christian Wiethe, J¨"org R¨"ommelt, Lukas
Networks on Jetson TX1. In Proceedings of the International Conference on Parallel Baur, Bj¨"orn H¨"ackel, and Alexander Sauer. 2022. How Sustainable is Machine
and Distributed Processing Techniques and Applications (PDPTA). The Steering Learning in Energy Applications? – The Sustainable Machine Learning Balance
Committee of The World Congress in Computer Science, 375–382. Sheet. https://aisel.aisnet.org/wi2022/sustainable_it/sustainable_it/1
[89] Friederike Rohde, Maike Gossen, Josephin Wagner, and Tilman Santarius. 2021. [109] Martina Willenbacher, Torsten Hornauer, and Volker Wohlgemuth. 2021. Re-
Sustainability challenges of Artificial Intelligence and Policy Implications. Ökol- bound Effects in Methods of Artificial Intelligence. In Advances and New
ogisches Wirtschaften-Fachzeitschrift 36, O1 (2021), 36–40. Trends in Environmental Informatics. Springer, Cham, Switzerland, 73–85. https:
[90] Bita Darvish Rouhani, Azalia Mirhoseini, and Farinaz Koushanfar. 2016. De- //doi.org/10.1007/978-3-030-88063-7_5
Light: Adding Energy Dimension To Deep Neural Networks. In ISLPED ’16: [110] Carole-Jean Wu, Ramya Raghavendra, Udit Gupta, Bilge Acun, Newsha Ardalani,
Proceedings of the 2016 International Symposium on Low Power Electronics and Kiwan Maeng, Gloria Chang, Fiona Aga, Jinshi Huang, Charles Bai, et al. 2022.
Design. Association for Computing Machinery, New York, NY, USA, 112–117. Sustainable AI: Environmental implications, challenges and opportunities. Pro-
https://doi.org/10.1145/2934583.2934599 ceedings of Machine Learning and Systems 4 (2022), 795–813.
[91] Kanokwan Rungsuptaweekoon, Vasaka Visoottiviseth, and Ryousei Takano. [111] Haichuan Yang, Yuhao Zhu, and Ji Liu. 2019. Energy-constrained compression
2017. Evaluating the power efficiency of deep learning inference on embedded for deep neural networks via weighted sparse projection and layer input masking.
GPU systems. In 2017 2nd International Conference on Information Technology International Conference on Learning Representations (ICLR) (2019).
(INCIT). IEEE, 1–5. [112] Tien-Ju Yang, Yu-Hsin Chen, and Vivienne Sze. 2017. Designing Energy-Efficient
[92] Roy Schwartz, Jesse Dodge, Noah A. Smith, and Oren Etzioni. 2020. Green AI. Convolutional Neural Networks Using Energy-Aware Pruning. In Proceedings
Commun. ACM 63, 12 (Nov. 2020), 54–63. https://doi.org/10.1145/3381831 of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR).
[93] Ilia Shumailov, Yiren Zhao, Daniel Bates, Nicolas Papernot, Robert Mullins, [113] Xiangyu Yang, Sheng Hua, Yuanming Shi, Hao Wang, Jun Zhang, and Khaled B.
and Ross Anderson. 2021. Sponge examples: Energy-latency attacks on neural Letaief. 2020. Sparse Optimization for Green Edge AI Inference. Journal of
networks. In 2021 IEEE European Symposium on Security and Privacy (EuroS&P). Communications and Information Networks 5, 1 (2020), 1–15. https://doi.org/10.
IEEE, 212–231. 23919/JCIN.2020.9055106
[94] Martino Sorbaro, Qian Liu, Massimo Bortone, and Sadique Sheik. 2020. Opti- [114] Zhaohui Yang, Mingzhe Chen, Walid Saad, Choong Seon Hong, and Mohammad
mizing the energy consumption of spiking neural networks for neuromorphic Shikh-Bahaei. 2020. Energy efficient federated learning over wireless communi-
applications. Frontiers in neuroscience 14 (2020), 662. cation networks. IEEE Transactions on Wireless Communications 20, 3 (2020),
[95] Dimitrios Stamoulis, Ermao Cai, Da-Cheng Juan, and Diana Marculescu. 2018. 1935–1949.
HyperPower: Power- and memory-constrained hyper-parameter optimization [115] Chunrong Yao, Wantao Liu, Weiqing Tang, Jinrong Guo, Songlin Hu, Yijun Lu,
for neural networks. In 2018 Design, Automation, and Test in Europe Conference. and Wei Jiang. 2021. Evaluating and analyzing the energy efficiency of CNN
19–24. https://doi.org/10.23919/DATE.2018.8341973 inference on high-performance GPU. Concurrency and Computation: Practice
[96] Dimitrios Stamoulis, Ting-Wu Rudy Chin, Anand Krishnan Prakash, Haocheng and Experience 33, 6 (2021), e6064.
Fang, Sribhuvan Sajja, Mitchell Bognar, and Diana Marculescu. 2018. Designing [116] Barzan A Yosuf, Sanaa H Mohamed, Mohammed M Alenazi, Taisir EH El-Gorashi,
adaptive neural networks for energy-constrained image classification. In 2018 and Jaafar MH Elmirghani. 2021. Energy-Efficient AI over a Virtualized Cloud
IEEE/ACM International Conference on Computer-Aided Design (ICCAD). ACM, Fog Network. In Proceedings of the Twelfth ACM International Conference on
1–8. Future Energy Systems. 328–334.
[97] Emma Strubell, Ananya Ganesh, and Andrew McCallum. 2019. Energy and [117] Jia-Ruei Yu, Chun-Hsien Chen, Tsung-Wei Huang, Jang-Jih Lu, Chia-Ru Chung,
Policy Considerations for Deep Learning in NLP. In Proceedings of the 57th Ting-Wei Lin, Min-Hsien Wu, Yi-Ju Tseng, and Hsin-Yao Wang. 2022. Energy
Annual Meeting of the Association for Computational Linguistics. Association for Efficiency of Inference Algorithms for Clinical Laboratory Data Sets: Green
Computational Linguistics, Florence, Italy, 3645–3650. https://doi.org/10.18653/ Artificial Intelligence Study. J. Med. Internet Res. 24, 1 (Jan. 2022), e28036.
v1/P19-1355 https://doi.org/10.2196/28036
[98] Yuyang Sun, Zhixin Ou, Juan Chen, Xinxin Qi, Yifei Guo, Shunzhe Cai, and [118] Boyu Zhang, Azadeh Davoodi, and Yu Hen Hu. 2018. Exploring Energy and
Xiaoming Yan. 2021. Evaluating Performance, Power and Energy of Deep Neural Accuracy Tradeoff in Structure Simplification of Trained Deep Neural Networks.
IEEE Journal on Emerging and Selected Topics in Circuits and Systems 8, 4 (2018), [120] Sha Zhu, Kaoru Ota, and Mianxiong Dong. 2021. Green AI for IIoT: Energy
836–848. https://doi.org/10.1109/JETCAS.2018.2833383 Efficient Intelligent Edge Computing for Industrial Internet of Things. IEEE
[119] Xingzhou Zhang, Yifan Wang, and Weisong Shi. 2018. pCAMP: Performance Trans. Green Commun. Networking 6, 1 (Aug. 2021), 79–88. https://doi.org/10.
Comparison of Machine Learning Packages on the Edges. In USENIX workshop 1109/TGCN.2021.3100622
on hot topics in edge computing (HotEdge 18).

A Systematic Review of Green AI: Roberto Verdecchia June Sallou Luís Cruz

Uploaded by

Copyright:

Available Formats

A Systematic Review of Green AI: Roberto Verdecchia June Sallou Luís Cruz

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

A Systematic Review of Green AI: Roberto Verdecchia June Sallou Luís Cruz

Uploaded by

Copyright:

Available Formats

A Systematic Review of Green AI

Roberto Verdecchia June Sallou Luís Cruz

perparameters to improve model sustainability, or benchmarking

Green AI studies, albeit most target academic readers. Green AI tool

1 Introduction system’s lifecycle (e.g, data collection, training, monitoring), dif-

Figure 1: Systematic literature review process overview.

• Intended reader: If the study is primarily intended for aca- 33

51 Deployment ⊲ 17 out of 98 papers. Addressing the deployment

Other ⊲ 5 out of 98 papers. Studies addressing a relevant topic Green AI domains

Figure 11: Occurrence of research strategies used in the

3 https://developer.nvidia.com/embedded-computing. Accessed 23th December 2022.

You might also like