Professional Documents
Culture Documents
Keywords: Continuous developments in data science have brought forth an exponential increase in complexity of
Sustainability machine learning models. Additionally, data scientists have become ubiquitous in the private market and
Carbon emission academic environments. All of these trends are on a steady rise, and are associated with an increase in
AI
power consumption and associated carbon footprint. The increasing carbon footprint of large-scale advanced
Data science
data science has already received attention, but the latter trend has not. This work aims to estimate the
Energy consumption
Carbon footprint
contribution of the increasingly popular ‘‘common’’ data science to the global carbon footprint. To this end,
Common data science the power consumption of several typical tasks in the aforementioned common data science tasks are measured
and compared to: large-scale ‘‘advanced’’ data science, common computer-related tasks, and everyday non-
computer related tasks. An automated data science project is also run on various hardware architectures.
To assess its sustainability in terms of carbon emission, the measurements are converted to gCO2 eq and an
equivalent unit of ‘‘km driven by car’’. Our main findings are: ‘‘common’’ data science consumes 2.57 more
power than regular computer usage, but less than some common everyday power-consuming tasks such as
lighting or heating; advanced data science consumes substantially more power than common data science,
and can be either on par or vastly surpass common everyday power-consuming tasks, depending on the scale
of the project. In addition to the reporting of these results, this work also aims to inspire researchers to include
power usage and estimated carbon emission as a secondary result in their work.
https://doi.org/10.1016/j.suscom.2023.100864
Received 31 May 2022; Received in revised form 6 January 2023; Accepted 20 February 2023
Available online 25 February 2023
2210-5379/© 2023 The Authors. Published by Elsevier Inc. This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-
nc-nd/4.0/).
B. Meulemeester and D. Martens Sustainable Computing: Informatics and Systems 38 (2023) 100864
and communication networks account for 25%; the sum of which Table 1
is comparable to the estimate of data center GHGe of 45%. The Overview of considered experiments showing benefits and drawbacks, showing
three criteria for each experiment: Hardware representativeness, workflow realism
contribution to the global GHGe of household ICT devices, such as those (how well the considered task mirrors a real-world task), and workflow repre-
used by common data scientists, should not be underestimated. sentativeness (compared to other data scientists). The experiment indices follow
This work aims to compare the previously mentioned advanced the enumeration as given in Section 2.
data science, or ‘‘big datacenter data science’’ to the latter common Experiment 1 2
data science. The main question is if the latter common data science Hardware Representative 1 architecture Several architectures
has a substantial contribution to the global GHGe. To this end, the Workflow Realistic Real workflow Automated
carbon emissions of common data science is estimated and compared to Representative Singular Singular & automated
common computer tasks, common non-computer related everyday tasks
and the aforementioned ‘‘big datacenter’’ advanced data science. This is
done by means of performing a common data science project and other (NCR) and advanced data science (ADS). During this cross-comparison,
various tasks on a laptop while measuring their power consumption, all computer related tasks and common data science tasks were per-
and comparing these values to one another. Additionally, an automated formed on the same hardware architecture. The considered hardware
data science project is run on various hardware architectures, and an architecture is described in Table 2 under index 3.
empirical law is constructed to extrapolate these results to even more
Only during this experiment is the hardware limited to a single
hardware architectures. Further details of these experiments are given
hardware structure. This is done in order to allow for consistent com-
in Section 2.
parison between these tasks; the aim of this experiment is to asses an
The goal of this work is not to report a large-scale benchmark on
estimate of the relative difference in power consumption of these tasks.
the power consumption and environmental impact of common data
science; such results are only possible and accurate if these numbers are (CR) Baseline computer (idle). This task is done by leaving a computer
present plentifully in the literature. This is the only way to accumulate running with the screen on, and no open programs. The value of this
realistic, trustworthy data that covers a broad range of data scientists, task varies depending on the efficiency of the hardware, background
data science projects and different hardware architectures. This work tasks and screen brightness.
does aim to aid the popularization of reporting of such numbers as
secondary results in order to make this a possibility. Such reporting (CR) Watching a movie. This task is completed by streaming Avengers:
is not yet standard practice, as is demonstrated by the lack of these Age of Ultron via Disney+ on Firefox 96.0 for Ubuntu. No other tabs
numbers in hyper-scale AI projects such as DALL⋅E [16] and GPT-3 [3], or programs were running. The carbon emission of the datacenter
but also in many other data science related works. providing the movie is not taken into account due to lack of available
data.
2. Methods (CR) Normal computer usage. Normal usage entails browsing the in-
ternet, reading a pdf and opening/closing various programs: Inkscape,
In order to assess the environmental impact of performing common Firefox, text editor and file browser. Tasks that involve video rendering
data science, two different experiments are performed: (such as gaming, watching Youtube or viewing local video files) are
1. Performing a real data science project on a single hardware purposely avoided, as these are comparable to the task Watching a
device and comparing it to other common computer related tasks movie.
on the same hardware device, common non-computer related (CR) Working in excel. This task entails making plots and performing
tasks, and advanced data science. various column transformations on a 4, 000 × 20 dataset containing
2. Automating the previous data science project and running it on only continuous numerical values. These column transformations aim
multiple devices, comparing the outcomes. to reflect common data science operations such as: deleting columns,
The goal of these experiments is to assess the contribution of common scaling columns, and power- and log-transformations. Plots are limited
data science to the overall power consumption of one or more devices. to histograms. These Excel operations are performed in LibreOffice Calc
This contribution in power consumption is then converted to carbon version 1:6.4.7-0ubuntu0.20.04.2.
emission as explained in Section 2.6. As it is not feasible to consider
(CDS) Data science project. This task entails doing a common data sci-
every possible combination of hardware architectures as individual
ence project, namely a credit scoring classification problem: identifying
parameters, and such large-scale benchmark is not the scope of this
defaulters on a loan. This project is split into two main parts:
work, we consider instead three essential properties that an experiment
must have in order to reflect an accurate assessment of the power 1. Exploration & preprocessing
consumption of common data science: 2. Gridsearch & fit
1. Hardware representativeness
Exploration & preprocessing entails constructing plots of features
2. Workflow representativeness
and processing features according to their meaning, type and distribu-
3. Workflow realism
tion. This is done by means of column transformations on a 20, 000 ×
How well each experiment covers these properties is shown in Table 1, 35 dataset with mixed continuous and categorical features. All data
and revisited in Section 4.5. Note that this work is inherently limited exploration and processing is done in Python3.8 using matplotlib,
by the workflow of the authors, and no experiment in this work can pandas and scikit-learn. Column transformations includes: encoding
perfectly simulate or represent the workflow of other data scientists. (one-hot encoding and WOE-encoding), scaling to normal distributions
More details on both is provided below. with QuantileTransformer(), filling missing values with means
or modes and over- and undersampling with ADASYN and Rando-
2.1. Experiment №1: Performing a real data science project and comparing mUnderSampler(). The exact workflow can be found in the auto-
to other common tasks mated version of the data science project (see Section 2.2), available
on github at [17].
This section provides a detailed overview of each task that was per- Gridsearch & fit entails performing two hyperparameter gridsearches
formed in order to compare common data science (CDS) to other com- for the machine learning models Random Forest, KNN and Logistic
mon computer related tasks (CR), common non-computer related tasks Regression on the dataset as described in the previous paragraph.
2
B. Meulemeester and D. Martens Sustainable Computing: Informatics and Systems 38 (2023) 100864
Table 2
Overview of all hardware types considered in this work. The indices are ordered to be congruent with the graphs.
Index Name CPU Threads RAM slots RAM
Dell Inc. Intel(R) Core(TM) SODIMM DDR4 Synchronous
3 Inspiron 7570 i7-8550U CPU 8 1x Unbuffered (Unregistered)
(07EA) Notebook @ 1.80 GHz 2400 MHz (0,4 ns)
Dell Inc. Intel(R) Core(TM) SODIMM DDR4 Synchronous
2 XPS 15 9560 i7-7700HQ CPU 8 2x Unbuffered (Unregistered)
(07BE) Notebook @ 2.80 GHz 2400 MHz (0,4 ns)
Dell Inc. Intel(R) Core(TM)
SODIMM DDR4 Synchronous
1 XPS 15 9500 i9-10885H CPU 16 2x
3200 MHz (0,3 ns)
(097D) Notebook @ 2.40 GHz
11th Gen
Dell Inc. SODIMM DDR4 Synchronous
Intel(R) Core(TM) 16 1x
0 Latitude 5421 3200 MHz (0.3 ns)
i7-11850H
(0A66) Notebook
@ 2.50 GHz
Table 3
Hyperparameter grids for the two gridsearches performed during the task Gridsearch & fit.
Model Gridsearch #1 Gridsearch #2
n_estimators: [100, 500, 1000, 2000] n_estimators: [2000, 2500, 3000]
Random Forest max_depth: [10, 50, 100] max_depth: [50, 70, 90]
min_samples_split: [2, 10] min_samples_split: [2]
n_neighbors: [10, 20, 50, 100, 500, 1000] n_neighbors: [10, 20, 50]
KNN
p: [1, 2] p: [1]
penalty: [l2, l1, elasticnet] penalty: [l2]
Logistic Regression C: [0.001, 0.01, 0.1, 1, 10, 100, 1000] C: [0.1, 0.3162, 1.0, 3.162, 10.0]
max_iter: [10, 100, 500, 1000] max_iter: [1500, 2000, 3000]
Gridsearches are performed by starting with a coarse Gridsearch and re- • This value was measured using different hardware.
fining the grid intervals once with a second Gridsearch. The considered • The dataset was already processed; processing does not contribute
hyperparameter grids are shown in Table 3. A 5-fold cross-validation to the reported value.
scheme is used. • Only one training session is reported, which does not reflect a
realistic case of trial and error, with multiple training sessions.
(NCR) Burning a lightbulb for an hour. Assuming a 10 W light bulb. • While computationally challenging, it is feasible to replicate this
task on a home computer. This measurement lives in the gray
(NCR) Streaming season 1 of friends. The carbon emission of this task
zone between common and advanced data science.
was calculated using the same measurement for power consumption as
the task Watching a movie and extrapolating this to the duration of the (ADS) Training GPT-3 in the EU. This measurement is an estimate
first season of Friends (2h28′ ). taken from [6], assuming Microsoft’s average datacenter PUE (Power
Usage Effectiveness) of 2015, re-calculated with updated values for the
(NCR) Leaving the office lights on over the weekend. This is calculated by GHGe associated with energy production (EU28 average from 2020)
assuming the office is lit by 8 10 W TL lights, burning from Friday 5pm and GHGe of cars (EU28 average from 2021).
until Monday 9am: the lighting setup at the office of the first author.
2.2. Experiment №2: Running an automated data science project on multi-
(NCR) Heating an office for a day. This is calculated by assuming an
ple hardware architectures
office with following properties:
3
B. Meulemeester and D. Martens Sustainable Computing: Informatics and Systems 38 (2023) 100864
3. Performing basic processing on the data to allow for training. 𝙽 > 𝚘𝚞𝚝𝚙𝚞𝚝.𝚝𝚡𝚝 (1)
4. Training a logistic regression model on unprocessed data.
5. Iteratively: where -D enables the -R option, -R denotes sampling should be done
on the RAPL interface, -f enables showing the average CPU frequency,
(a) Improved processing on all features. -d denotes a delay before starting the measurement (in seconds), 𝛥𝑡
(b) Checking effect of previous processing by re-training the denotes the time interval between samples (in seconds), N denotes
logistic regression model. the amount of samples and -g enables measuring the GPU power
(c) Plotting distributions of processed features. consumption as well. GPU power consumption and GPU programming
are not explored in this work, as the authors consider this advanced
6. Performing a coarse gridsearch.
data science, rather than common data science. The sampling interval
7. Performing a finer gridsearch.
should be sufficiently small compared to the duration of the task in
The code for this automated workflow is available on github at [17]. order to accurately measure the distribution of its power usage.
While the code allows for multiple iterations of improving data pro- Where possible, CarbonTracker [6] is also used to measure energy
cessing and testing the effect of these improved processing steps on the consumption. CarbonTracker is designed to track and predict the en-
resulting AUC score with a logistic regression model, only one iteration ergy consumption of training AIs, but can be used for any piece of code.
is actually performed. Multiple iterations are not included in this work, It must be noted that CarbonTracker was designed for measuring the
since any choice of number for the amount of iterations would be power consumption of training an AI during one or more epochs, and
arbitrary, and would not increase the realism of the automated data
not shorter code snippets prevalent in common data science. Due to
science project by any measurable metric. Additionally, the authors
this, its accuracy may deviate, especially if the runtime of the code is
argue that a singular iteration of data processing improvement is not
short. For this reason, interpreting the data will be based on the RAPL
too unrealistic for a data science project, albeit a minimum. However,
measurements, and the CarbonTracker measurements are only shown
this argument is subject to the specificity of the author’s workflow, as
for completeness, and as a benchmark to check if it produces similar
mentioned before at the end of Section 2 and Table 1, and is open to
results. Deviating results are also discussed in Section 4.
adversity.
Inevitably, assumptions need to be made in order to compare the As mentioned before, both sampling the RAPL interface directly,
measurements in a comprehensive manner. The measurements in this as well as CarbonTracker [6] is used to measure the power/energy
work are only as truthful as the truth value of these assumptions. These
consumption. The resulting values should be interpreted differently.
will also be revisited in Section 4.5.
The RAPL interface. This method measures by directly sampling the
1. It suffices to measure the power consumption of the CPU to gain
power consumption of the CPU every other time interval. This mea-
insight in the total power consumption of performing various
surement is code-independent. Whether or not your code is running,
computer tasks, as we can estimate the total power consumption
this method measures the power consumption. This method gives a
by assuming a CPU to RAM power consumption of 3 ∶ 2, as
good overview of the entire power consumption of some task, including
estimated by Anthony et al. [6].
2. Differences in power consumption of the automated data science the non-code related parts, e.g. looking up documentation and articles,
project across the measured hardware architectures are repre- debugging, running code that finishes with an error, etc. This mea-
sentative for differences in power consumption of data science surement gives the most accurate measurement of a certain task, but
projects for other hardware architectures. will include variability depending on the efficiency, experience and
3. The energy consumption of common data science on other hard- knowledge of the data scientist performing the task. Note that only
ware structures can, to some extent, be estimated based using the power consumption of the CPU is considered. This may deviate
the Thermal Design Power (TDP) as a proxy for the power from the true power consumption, especially when running code that
consumption of the entire machine, and assuming an inverse relies heavily on RAM. This drawback is discussed in Section 4.5, when
relationship between the total consumed energy for a single revisiting the assumptions made earlier (see Section 2.3).
automated data science task 𝐸𝐴𝑢𝑡𝑜𝐷𝑆 and TDP of the nature:
CarbonTracker. This method measures the energy consumption of a
𝐸𝐴𝑢𝑡𝑜𝐷𝑆 = 𝑎∕𝑇 𝐷𝑃 + 𝑏, where 𝑎 and 𝑏, the dropoff rate and
single piece of code. This measurement gives a good overview of how
intercept respectively, are parameters that are to be estimated.
much energy was consumed by running the code, and nothing but the
4. A car emits as much gCO2 eq as the EU average from 2019.
code. If a script finishes on an error, no measurement is written out.
5. Energy production emits as much CO2 eq as the EU average from
2020. This measurement has a higher repeatability, but lacks information
on the overhead of the tasks: the non-code related parts and running
2.4. Measuring the power: Measurement tools bugged code that finishes with an error.
4
B. Meulemeester and D. Martens Sustainable Computing: Informatics and Systems 38 (2023) 100864
Fig. 1. Power consumption distributions of working on common computer tasks and Fig. 2. Energy consumption and carbon emission (in km driven by car equivalent) of
data science, measured by CarbonTracker (dark blue) and sampling the RAPL interface working on common computer tasks and data science related activities. Only RAPL
(light teal). RAPL samples are taken every 10 s, while CarbonTracker samples are taken measurements are shown.
every successful run of a piece of code. (For interpretation of the references to color
in this figure legend, the reader is referred to the web version of this article.)
2.6. Converting
where CO2 eq denotes the amounts of gram CO2 that would yield a
greenhouse effect of equal magnitude as the considered greenhouse
gases, i.e. the seven greenhouse gases considered by the Kyoto Proto-
col [22]: carbon dioxide (CO2 ), methane (NH4 ), nitrous oxide (N2 O),
hydrofluorocarbons (HFCs), perfluorocarbons (PFCs), sulphur hexaflu-
oride (SF6 ) and nitrogen trifluoride (NF3 ).
In order to increase the interpretability of the gCO2 eq quantity,
they are converted to the quantity ‘‘km driven by car’’, as suggested
by Anthony et al. [6]. To do so, we use the average GHGe of every
registered car in the EU up until 2019 [23], yielding a distance per
equivalent carbon emission of:
3. Results
Fig. 3. Carbon emission of common non-computer related tasks and data science
3.1. Experiment №1: Performing a real data science project and comparing (common and large-scale) on a logarithmic scale, expressed in units of km driven by car.
to other common tasks Values for common data science and computer-related tasks are reported for hardware
type 3 (see Table 2).
5
B. Meulemeester and D. Martens Sustainable Computing: Informatics and Systems 38 (2023) 100864
Fig. 4. Power distributions for running an automated data science project on different Fig. 5. Total energy used for running an automated data science project on different
hardware structures as described in Section 2.2. The experiment index and CPU are hardware structures as described in Section 2.2. The experiment index and CPU are
given as axis labels, where the indices are congruent with the hardware as shown in given as axis labels, where the indices are congruent with the hardware as shown in
Table 2. Table 2.
3.4. Comparison of considered CPUs to other CPUs Power consumption. We can see in Fig. 1 how CarbonTracker yields
similar results on power consumption as directly sampling the RAPL
interface for Gridsearch & fit, but there is a notable difference between
The Thermal Design Power (TDP) of all CPUs recorded by Passmark the two measurements for the task Exploration & preprocessing. This can
[20] are shown in Fig. 7. The TDP of the CPU considered in Experiment be explained by the fact that the latter task includes a lot of overhead,
№1 is denoted with a blue line, while all additional CPUs considered by such as looking up documentation, scaling columns one by one, pro-
Experiment №2 have a TDP equal to 45 W, as denoted by a gold line. gramming, debugging and running faulty code. All of the previously
6
B. Meulemeester and D. Martens Sustainable Computing: Informatics and Systems 38 (2023) 100864
7
B. Meulemeester and D. Martens Sustainable Computing: Informatics and Systems 38 (2023) 100864
8
B. Meulemeester and D. Martens Sustainable Computing: Informatics and Systems 38 (2023) 100864
While the uprising in data science jobs can indeed be associated with [4] A. Rizwan, A. Zoha, I.B. Mabrouk, H.M. Sabbour, A.S. Al-Sumaiti, A. Alomainy,
an increased carbon footprint, it will prove much more substantial M.A. Imran, Q.H. Abbasi, A review on the state of the art in atrial fibrillation
detection enabled by machine learning, IEEE Rev. Biomed. Eng. 14 (2020)
to be mindful about the power consumption and associated carbon
219–239.
emission of some everyday tasks. Efficient heaters and lighting, insu- [5] M. Holmstrom, D. Liu, C. Vo, Machine learning applied to weather forecasting,
lation, green energy production and green travel methods will always Meteorol. Appl. 10 (2016) 1–5.
beat, by a landslide, preferring e.g. a RandomsearchCV() over a [6] L.F.W. Anthony, B. Kanding, R. Selvan, Carbontracker: Tracking and predicting
GridsearchCV() when doing common data science. the carbon footprint of training deep learning models, 2020, arXiv:2007.03051.
[7] G. Cook, J. Lee, T. Tsai, A. Kongn, J. Deans, B. Johnson, E. Jardim, B. Johnson,
That does not mean, however, that efforts to reduce the carbon Clicking Clean: Who Is Winning the Race to Build a Green Internet? Technical
footprint of common data science are in vain. Mindfulness of the report, Greenpeace, 2017.
environmental impact of data science remains important, whether it [8] E. Masanet, A. Shehabi, N. Lei, S. Smith, J. Koomey, Recalibrating global
is applied to common or large-scale data science. The methods de- data center energy-use estimates, Science 367 (6481) (2020) 984–986, http:
//dx.doi.org/10.1126/science.aba3758.
scribed in this work can be used for both. Such mindfulness has not
[9] Z. Cao, X. Zhou, H. Hu, Z. Wang, Y. Wen, Towards a systematic survey for
been reached, given the absence of such numbers in hyper-scale data carbon neutral data centers, 2022, arXiv:2110.09284.
science projects such as GPT-3 and DALL⋅E, but also in most other [10] L. Belkhir, A. Elmeligi, Assessing ICT global emissions footprint: Trends
data science related works. If the reporting of the carbon footprint to 2040 & recommendations, J. Clean. Prod. 177 (2018) 448–463, http://
of data science becomes ubiquitous, independent of the scale of the dx.doi.org/10.1016/j.jclepro.2017.12.239, URL: https://www.sciencedirect.com/
science/article/pii/S095965261733233X.
data science project, it would unlock the possibility of assessing the [11] IEA, Global data centre energy demand by data centre type, 2010–2022,
contribution of common and advanced data science to global GHGe to 2021, https://www.iea.org/data-and-statistics/charts/global-data-centre-energy-
a much more precise and realistic degree. demand-by-data-centre-type-2010-2022. (Last accessed 08 March 2022).
[12] R. Bashroush, A. Lawrence, Beyond PUE: Tackling IT’s Wasted Terawatts, Uptime
Institute, 2020, Available from: https://uptimeinstitute.com/beyond-puetackling-
CRediT authorship contribution statement
it’s-wasted-terawatts. (Accessed 11 June 2020).
[13] N. Rteil, R. Bashroush, R. Kenny, A. Wynne, Interact: IT infrastructure energy
Bjorge Meulemeester: Conceptualization, Data curation, Formal and cost analyzer tool for data centers, Sustain. Comput. Inform. Syst. 33
analysis, Investigation, Methodology, Resources, Software, Validation, (2022) 100618, http://dx.doi.org/10.1016/j.suscom.2021.100618, URL: https:
Visualization, Roles/Writing – original draft, Writing – review & edit- //www.sciencedirect.com/science/article/pii/S2210537921001062.
[14] News LinkedIn, LinkedIn jobs on the rise 2022: The 25 U.S. roles that are
ing. David Martens: Conceptualization, Funding acquisition, Investiga-
growing in demand, 2022, https://www.linkedin.com/pulse/linkedin-jobs-rise-
tion, Methodology, Project administration, Supervision, Roles/Writing 2022-25-us-roles-growing-demand-linkedin-news. (Last accessed 28 February
– original draft, Writing – review & editing. 2022).
[15] News LinkedIn, LinkedIn jobs on the rise 2022: The roles that are grow-
ing in demand, 2022, https://www.linkedin.com/pulse/linkedin-jobs-rise-2022-
Declaration of competing interest
roles-growing-demand-linkedin-news-europe/. (Last accessed 28 February 2022).
[16] A. Ramesh, M. Pavlov, G. Goh, S. Gray, C. Voss, A. Radford, M. Chen, I.
The authors declare that they have no known competing finan- Sutskever, Zero-shot text-to-image generation, 2021, http://dx.doi.org/10.48550/
cial interests or personal relationships that could have appeared to ARXIV.2102.12092, URL: https://arxiv.org/abs/2102.12092.
influence the work reported in this paper. [17] B. Meulemeester, D. Martens, How sustainable is ‘‘common’’ data science
in terms of power consumption? 2022, URL: https://github.com/bgmeulem/
EmissionCommonDS.
Data availability [18] electricpoint, Electric heating room size calculator, 2022, https://www.
electricpoint.com/heating/electric-heating/how-to-calculate-kw-required-to-
How the data on energy usage is obtained is described in de- heat-a-room. (Last accessed 01 March 2022).
[19] ICAO, ICAO carbon emissions calculator, 2016, https://www.icao.int/
tail. We encourage readers to report the same numbers on their own
environmental-protection/CarbonOffset/Pages/default.aspx. (Last accessed
architecture and data science project. 01 March 2022).
[20] Passmark, CPU mega list, 2022, https://www.cpubenchmark.net/CPU_mega_
References page.html. (Last accessed 01 March 2022).
[21] European Environment Agency EEA, Greenhouse gas emission intensity of
[1] M. Hiransha, E.A. Gopalakrishnan, V.K. Menon, K. Soman, NSE stock mar- electricity generation by country, 2021, https://www.eea.europa.eu/data-and-
ket prediction using deep-learning models, Procedia Comput. Sci. 132 (2018) maps/daviz/co2-emission-intensity-9. (Last accessed 04 February 2022).
1351–1362. [22] K. Protocol, Kyoto Protocol, UNFCCC Website, 1997, Available Online: http:
[2] A. Toth, L. Tan, G. Di Fabbrizio, A. Datta, Predicting shopping behavior with //unfccc.int/kyoto_protocol/items/2830.php. (Accessed 1 January 2011).
mixture of RNNs, 2017, ecom@ sigir. [23] European Environment Agency EEA, CO2 performance of new passenger
[3] T.B. Brown, B. Mann, N. Ryder, M. Subbiah, J. Kaplan, P. Dhariwal, A. cars in Europe, 2021, https://www.eea.europa.eu/ims/co2-performance-of-new-
Neelakantan, P. Shyam, G. Sastry, A. Askell, S. Agarwal, A. Herbert-Voss, G. passenger. (Last accessed 04 February 2022).
Krueger, T. Henighan, R. Child, A. Ramesh, D.M. Ziegler, J. Wu, C. Winter, C. [24] A. Menezes, A. Cripps, R. Buswell, J. Wright, D. Bouchlaghem, Estimating
Hesse, M. Chen, E. Sigler, M. Litwin, S. Gray, B. Chess, J. Clark, C. Berner, the energy consumption and power demand of small power equipment in
S. McCandlish, A. Radford, I. Sutskever, D. Amodei, Language models are few- office buildings, Energy Build. 75 (2014) 199–209, http://dx.doi.org/10.1016/
shot learners, 2020, http://dx.doi.org/10.48550/ARXIV.2005.14165, URL: https: j.enbuild.2014.02.011, URL: https://www.sciencedirect.com/science/article/pii/
//arxiv.org/abs/2005.14165. S0378778814001224.
[25] I. Cutress, Why intel processors draw more power than expected: TDP
and turbo explained, 2018, https://www.anandtech.com/show/13544/why-intel-
processors-draw-more-power-than-expected-tdp-turbo. (Last accessed 08 March
2022).