The Energy Intensity of the Internet: Home and Access Networks

Abstract

Estimates of the energy intensity of the Internet diverge by several orders of magnitude. We present existing assessments and identify diverging de�finitions of the system boundary as the main reason for this large spread. The decision of whether or not to include end devices influences the result by 1-2 orders of magnitude. If end devices are excluded, customer premises equipment (CPE) and access networks havea dominant influence. Ofless influence is the consideration of cooling equipment and other overhead, redundancy equipment, and the amplifi�ers in the optical �bers. We argue against the inclusion of end devices when assessing the energy intensity of the Internet, but in favor of including CPE, access networks, redundancy equipment, cooling and other overhead as well as optical fi�bers. We further show that the intensities of the metro and core network are best modeled as energy per data, while the intensity of CPE and access networks are best modeled as energy per
time (i.e., power), making overall assessments challenging. The chapter concludes with a formula for the energy intensity of CPE and access networks. The formula is presented both in generic form as well as with concrete estimates of the average case to be used in quick assessments by practitioners. The following chapter develops a similar formula for the core and edge networks. Taken together, the two chapters provide an assessment method of the Internet's energy intensity that takes into account diff�erent modeling paradigms for di�fferent parts of the network.

Full text

The Energy Intensity of the Internet:
Home and Access Networks
Vlad C. Coroama1, Daniel Schien2, Chris Preist2, and Lorenz M. Hilty3,4,5
1Measure-IT Research, Bucharest, Romania
vlad.coroama@measureit-research.eu
2Department of Computer Science, University of Bristol, UK
3Department of Informatics, University of Zurich, Switzerland
4Empa, Swiss Federal Laboratories for Materials Science and Technology,
St. Gallen, Switzerland
5Centre for Sustainable Communications CESC, KTH Royal Institute of
Technology, Stockholm, Sweden
Abstract. Estimates of the energy intensity of the Internet diverge by
several orders of magnitude. We present existing assessments and iden-
tify diverging definitions of the system boundary as the main reason for
this large spread. The decision of whether or not to include end devices
influences the result by 1-2 orders of magnitude. If end devices are ex-
cluded, customer premises equipment (CPE) and access networks have
a dominant influence. Of less influence is the consideration of cooling
equipment and other overhead, redundancy equipment, and the ampli-
fiers in the optical fibers. We argue against the inclusion of end devices
when assessing the energy intensity of the Internet, but in favor of in-
cluding CPE, access networks, redundancy equipment, cooling and other
overhead as well as optical fibers. We further show that the intensities of
the metro and core network are best modeled as energy per data, while
the intensity of CPE and access networks are best modeled as energy per
time (i.e., power), making overall assessments challenging. The chapter
concludes with a formula for the energy intensity of CPE and access
networks. The formula is presented both in generic form as well as with
concrete estimates of the average case to be used in quick assessments
by practitioners. The following chapter develops a similar formula for
the core and edge networks. Taken together, the two chapters provide
an assessment method of the Internet’s energy intensity that takes into
account different modeling paradigms for different parts of the network.
Keywords: Internet, Energy Intensity, Energy Efficiency, Customer Premises
Equipment, Access Network
1 Introduction
Information and Communication Technologies (ICT) are increasingly perceived
as enablers of a reduction of anthropogenic greenhouse gas (GHG) emissions.
Studies providing evidence for this enabling effect come not only from academia
This Accepted Author Manuscript is copyrighted by Springer. The final publication will be
available via http://link.springer.com/bookseries/11156 by end of August 2014. Suggested citation:
Coroama, V.C., Schien, D., Preist, C., Hilty, L.M.: The Energy Intensity of the Internet: Home and
Access Networks. In: Hilty, L.M., Aebischer, B. (eds.) ICT Innovations for Sustainability. Advances
in Intelligent Systems and Computing 310, Springer International Publishing (2014, in press)
This Accepted Author Manuscript is copyrighted by Springer. The final publication will be
available via http://link.springer.com/bookseries/11156 by end of August 2014. Suggested citation:
Coroama, V.C., Schien, D., Preist, C., Hilty, L.M.: The Energy Intensity of the Internet: Home and
Access Networks. In: Hilty, L.M., Aebischer, B. (eds.) ICT Innovations for Sustainability. Advances
in Intelligent Systems and Computing 310, Springer International Publishing (2014, in press)
This Accepted Author Manuscript is copyrighted by Springer. The final publication will be
available via http://link.springer.com/bookseries/11156 by end of August 2014. Suggested citation:
Coroama, V.C., Schien, D., Preist, C., Hilty, L.M.: The Energy Intensity of the Internet: Home and
Access Networks. In: Hilty, L.M., Aebischer, B. (eds.) ICT Innovations for Sustainability. Advances
in Intelligent Systems and Computing 310, Springer International Publishing (2014, in press)

2V.C. Coroama, D. Schien, C. Preist, and L.M. Hilty
[1–3], but also from organizations as diverse as ICT industry associations [4],
the European Commission [5], and the World Wildlife Fund [6]. Reductions are
usually estimated based on quantitative scenarios [7], yielding an abatement
potential which has to be adjusted downwards by an estimate of the ICT appli-
cations own footprint to calculate the net effect.
ICT can reduce energy consumption and related GHG emissions through
three mechanisms: i) from optimization effects in domains such as smart en-
gines, buildings, or logistics, ii) due to ICT-supported novel paradigms for the
generation and distribution of electricity (i.e., smart grids), and iii) due to sub-
stitution effects in which information and communication services partly replace
other more energy-intensive activities [4,8].
Two problems must be solved to quantify the net effect of ICT applications
in these cases: First, the energy savings induced by ICT must be assessed. As we
argue in [9] and [10], this is methodologically challenging: The baseline scenario,
among other factors, as it expands into the future, is inherently speculative.
Moreover, allocation issues are raised by the fact that ICT typically does not
induce efficiency on its own, but only in a suitable technological, political, or
organizational context. Secondly, the energy consumption of the ICT solution
involved must be determined. This is also technically challenging, and existing
literature reports diverse results. The current and the subsequent chapter explore
this issue, with a particular focus on Internet services.
The energy intensity of the Internet, expressed as energy consumed to trans-
mit a given volume of data, is one of the most controversial issues. Existing
studies of the Internet energy intensity give results ranging from 136 kWh/GB
[11] down to 0.0064 kWh/GB [12], a factor of more than 20,000. Whether and to
what extent it is more energy efficient to download a movie rather than buying
the DVD, for example, or more sustainable to meet via videoconference instead
of travelling to a face-to-face meeting are questions that cannot be satisfyingly
answered with such diverging estimates of the substitute’s impact.
The Internet’s energy consumption, and the energy intensity of the Internet
as a suitable metric, are the topics of both the current chapter and the next
one. The current chapter presents a review of existing studies and provides ex-
planations for the large spread in their results. The chapter then recommends a
definition of the system boundary that is most useful for decision-making and
concludes by zooming in on the peripheral parts of the network and assessing
their contribution to the energy intensity of the Internet. The next chapter [13]
will focus on the remaining components, i.e., the core of the Internet.
2 Definitions and System Boundaries
In the late 1960s and early 1970s, the term Internet literally indicated the in-
terconnection of a small number of local area networks on university campuses;
back then, “the Internet” comprised a few routers and cables. Today, the Inter-
net is the vast and heterogeneous infrastructure connecting billions of computers
worldwide using the TCP/IP family of communication protocols. The majority

The Energy Intensity of the Internet: Home and Access Networks 3
of these computers belong to private users who connect to the Internet through
their Internet service provider (ISP).
Figure 1 presents a high-level structure of the Internet. We comment only
briefly on this structure here; more detailed discussions can be found in the
following chapter in section 2.1, and in the network models presented by [14,
15]. Users’ devices (such as desktop, laptop or tablet PCs and smartphones)
connect through what is referred to as customer premises equipment (CPE),
which are mainly WiFi routers and modems, to their ISP. The ISP bundles the
data from several users in multiplexers. These vary depending on the subscription
technology; for the widely used DSL connectivity they are digital subscriber
line access multiplexers (DSLAM). Together with the cables connecting them to
the CPE, these multiplexers constitute what is called the access network. After
passing through an edge router, the traffic enters the metro and core parts of
the Internet where routers with increasing capacities bundle the traffic. On the
other side, the traffic is decomposed according to its destination; a large part
is directed to data centers, while a smaller part is directed to other users (not
represented in Figure 1).
λ
CPE Metro Long Haul
Router Router
Access
Metro
Undersea
Data Center
λ
λ
λ
λ
λλ
λ
λ
Next ChapterThis Chapter
Provider
Edge
Fiber Optic
Transmission
Ring Topology Fiber Optic
Transmission
Mesh Topology
Fig. 1. Model of the Internet structure.
First attempts to understand the energy consumption of this distributed
and heterogeneous power-consumption system were undertaken a decade ago, in
2003-2004. Starting from statistical data and studies on the ICT equipment in
use, both [16] and [11] estimated the yearly power consumption of the Internet
in the US. Dividing this value by an estimate of the US Internet traffic for that
year resulted in estimates of the energy intensity of the Internet, i.e., the energy
consumed throughout the Internet per amount of data transferred.
The two studies differed in their definition of what constitutes the Internet.
While [16] considered only networking equipment (i.e., Ethernet hubs, LAN and
WAN switches in offices and buildings, together with the routers of the Internet),
[11] also took into account devices in data centers such as servers or data storage
(see Figure 1). As we discuss in more detail later, this definitional discrepancy

4V.C. Coroama, D. Schien, C. Preist, and L.M. Hilty
across studies persists until today, and is one of the main causes for the large
spread of the published results.
Before going into detail, we note that we discuss estimates for direct energy
consumption in the form of electricity only. The energy supply chain (containing
the supply of primary energy, power plants transforming primary energy sources
to electricity, and grids bringing them to the consuming devices) is excluded
from the system under study. We also exclude the “grey” energy embedded in
ICT hardware, although the material flows caused by producing and disposing
of hardware are significant (see [17] and the chapter by Hischier et al. [18] later
in this volume).
All studies that will be discussed apply an average allocation rule, distribut-
ing the equipment energy consumption evenly among the total traffic volume
over a certain period of time. We found no study focusing on marginal instead
of average effects.
3 Existing Assessments and Methods Used
The two assessments mentioned above, [11] and [16], both used the same method-
ology: dividing an estimate of the overall US Internet energy consumption by
the estimated total Internet traffic in the US. Since then, several more studies
used the same approach, while other studies deployed different methodologies.
According to the basic methodological approach they use, existing studies can
be classified in two classes:
Top-down. According to [19], top-down analyses are based on two estimates:
1) the overall energy demand of either the entire Internet or a part of it (e.g., a
country or a continent), and 2) the total Internet traffic of that region.
Bottom-up. By contrast, bottom-up approaches model parts of the Internet
(i.e., deployed number of devices of each type) based on network design princi-
ples. Such a model combined with manufacturers’ consumption data on typical
network equipment leads to an estimate of the overall energy consumption [15],
which is then related to an estimate of the corresponding data traffic. Bottom-
up studies can also use direct observations made in one or more case studies.
These provide primary data on one or more of the following: deployed equipment,
network topology and routing, power consumption of specific devices, and data
volume passing through specific devices. Case studies present energy intensity
values for specific cases, typically followed by a discussion of how the results can
be generalized.
3.1 Top-Down Assessments
The earliest top-down assessments were introduced above: Gupta et al. published
their results in 2003 [16], and Koomey et al. in 2004 [11]. The two studies used

The Energy Intensity of the Internet: Home and Access Networks 5
the same statistical inventory data on ICT equipment in commercial buildings
in the US for the year 2000 [20]. Despite building on the same inventory, they
make different assumptions as to which devices belong to the Internet and thus
yield distinct consumption results.
The assessment by Gupta et al. [16], taking into account Internet routers,
WAN and LAN switches as well as Ethernet hubs, yields a yearly energy con-
sumption of 6.05 terawatt-hours per year (TWh/a) for all networking devices
in the US. In assessing the energy intensity of “the Internet” in the US, the
study narrows the focus even more. It leaves campus devices (i.e., LAN switches
and hubs) outside the calculation, and considers only the consumption of WAN
switches and Internet routers, estimated at just 1.25 TWh/a in 2000.
Koomey et al. [11], on the other hand, consider not only the campus net-
working devices, but also data center devices – servers (10.2 TWh/a) and data
storage (1.5 TWh/a) – as well as uninterruptible power supplies (5.8 TWh/a),
leading to a total of 23.65 TWh/a. Furthermore, they multiply this result by 2
to account for overhead such as cooling and ventilation, leading to a total of 47
TWh/a, 37 times higher than the value in [16].
The two studies also use an identical source to estimate the US Internet
traffic: data from the Minnesota Internet Traffic Studies (MINTS) by Andrew
Odlyzko and colleagues. These data, published e.g. in [21], estimate a US Internet
traffic in 2000 of 20,000-35,000 TB/month. Using the same traffic data, and
consumption data different by a factor of 37, one could expect from the two
studies energy intensity results differing by the same factor 37. This, however,
is not the case. Koomey et al. [11] use the lower end of the Internet traffic
data from [21] but complement it with traffic on other public data networks
and private lines, leading to a total traffic of 348,000 terabyte per year (TB/a).
The study thus yields an energy intensity of 136 kilowatt-hours per Gigabyte
(kWh/GB) [11]. As for the other study, we can only speculate that Gupta et al.
[16] misinterpreted the data in [21] as yearly instead of monthly traffic values.
The study thus calculates with the range of 20,000-35,000 TB/a, which leads to
an energy intensity of 0.128-0.225 Joule/Byte, or 38-67 kWh/GB. The correct
energy intensity for the system boundaries used by [16], however, should have
been twelve times lower because the yearly traffic estimate was actually twelve
times higher. The corrected values for [16] are 3.2-5.6 kWh/GB (see Table 1).
An update for Koomey et al. [11] was published a few years later. The new
assessment by Taylor and Koomey [22] referred to the year 2006. Estimating
again the US Internet energy consumption and using three existing estimates of
the US Internet traffic per year, the new study yielded as result the range 8.8-
24.3 kWh/GB [22]. This 2006 estimate was yet again updated for the year 2008
in an article by Weber et al. [23]. For this period, the authors assumed that total
Internet traffic increased by 50% per year, and that total Internet electricity use
grew at a yearly rate of 14%, which had been the average global growth rate of
data center electricity use between 2000 and 2005. These assumptions resulted
in an average Internet electricity intensity of about 7 kWh/GB for 2008 [23].

6V.C. Coroama, D. Schien, C. Preist, and L.M. Hilty
The study by Lanzisera et al. [24] is another well-known top-down estimate.
The analysis only includes networking equipment, excluding not only end de-
vices but also the transmission lines. Estimating the total of both the US and
the world networking equipment stock for 2008, the power of each device and
their individual usage patterns, the article computed an annual electricity con-
sumption of 18 TWh for all networking equipment in the US and of 50.8 TWh
for the world. The study did not relate this consumption to traffic values to com-
pute the Internet energy intensity. To make the result comparable with other
studies, we divide it by an estimate for Internet data traffic for 2008 in order to
calculate the energy intensity. According to Cisco’s “Visual Networking Index,”
“global IP traffic grew 45 percent during 2009 to reach an annual run rate of 176
exabytes per year” [25]. We therefore assume a traffic volume of 121 exabytes
(EB) for 2008. Using this value as a worldwide traffic estimate for 2008 yields
an energy intensity of 0.39 kWh/GB for the world average.
3.2 Bottom-Up Assessments
The model-based approach has been used by Kerry Hinton’s research group at
the University of Melbourne [12, 14, 15,26, 27], as well as in [28] and [29]. Some of
these studies are not directly comparable to the results of top-down assessments
because they have different focuses such as analyzing only a part of the Internet
transmission (e.g., [27]) or analyzing the Internet power consumption per sub-
scriber and not per amount of data [26]. A few of these results may, nevertheless,
be adapted to be made compatible with studies on Internet energy intensity. As
we present in detail in [30], the very first assessment from the Melbourne group
[26] yields an Internet energy intensity of 0.91-2.52 kWh/GB, depending on the
estimate of worldwide Internet traffic used.
[14] provides a direct estimate of the energy intensity of Internet data trans-
mission: 75 micro-Joule per bit (µJ/bit), equal to 0.179 kWh/GB, at the access
rates typical of 2008. As the authors point out, their result represents a lower
bound or optimistic estimate in terms of energy consumption, because the model
assumes only state-of-the-art equipment and ignores the fact that less energy-
efficient legacy network equipment is still in use. They further state that they
expect this energy intensity to drop in the near future to 2-4 µJ/bit with in-
creasing access rates.
[12] puts forward a value of 2.7 µJ/bit. This value corresponds to 0.0064
kWh/GB and represents the lowest value published thus far. This study, how-
ever, aimed to compare the energy demand of traditional computing with that
of cloud computing. As the energy consumption of the access network is largely
independent of the traffic and would leave the result untouched, the authors
legitimately ignored it: “The access network does not influence the comparison
between conventional computing and cloud computing. Therefore, it is omitted
from consideration and is not included in our calculations of energy consump-
tion” [12]. While this assumption stands to reason within the scope of the study,
it can lead to misinterpretation if taken out of context, as we will discuss in the
next section.

The Energy Intensity of the Internet: Home and Access Networks 7
Finally, Schien et al. [29] used a network model to analyze the download of
the UK newspaper “The Guardian,” as well as the download of a 640 second
video from the Guardian’s video section. The newspaper’s html homepage was
located on a server within the UK, while video and images were outsourced
to a Content Distribution Network (CDN) and mirrored on several continents
within the CDN’s network. Downloads from clients in Oceania, North America,
and Europe were studied. Results showed that because of the CDN architecture,
geographical distance played only a minor role; the energy intensities of the
downloads from different continents were similar. For both the homepage and
the video, the intensity was 8-9 Joules per megabit (J/mbit), which corresponds
to approximately 0.02 kWh/GB. The study, while considering access, metro, and
core parts of the Internet, did not account for the CPE.
In [31], this work was extended to include CPE and end devices, and to
explore uncertainty and variability in assessments of digital services. In contrast
to this chapter and to [30], which try to represent the existing variability in
previous assessments, [31] estimates how uncertainty in energy intensity affects
the overall result. Combining earlier results in a triangular distribution, the
study arrived at a mean energy intensity for metro and core networks of 0.038
kWh/GB. For access network and CPE together, and excluding end devices, it
provided a mean energy estimate of 0.019 kWh/GB.
In [32], Coroam˘a and Hilty present an assessment of a 40 megabit per second
(Mbps) videoconferencing transmission of the case study introduced in [33]. For
a system boundary that included network devices and optical fibers but no end
devices, and making pessimistic assumptions in terms of energy consumption
where specific data were not available, the study yielded an energy intensity of
0.2 kWh/GB for 2009. As we argued in [32], many characteristics of the study
(such as an above-average number of hops) justify considering its result above-
average in terms of energy intensity. This implies that the case-study result,
when generalized, should be considered an upper bound for the average energy
intensity.
The setting of this case study was such that no CPE or access network was
distinguishable. The conference was held between a large conference center in
Switzerland and a university campus in Japan. Both sites were directly connected
to the metro network in the same way that Figure 1 depicts data centers. Yet,
the edge routers on each side of the connection behaved similarly as CPE and
the access network behave in the typical setting: they had a low load far from
their capacity, and with an energy consumption that had to be allocated entirely
or to a large extent to the case study.
4 Factors Influencing the Results
Results of the surveyed studies span a very wide range: from the 136 kWh/GB
of [11] down to the 0.006 kWh/GB from [12], there is a spread of four orders
of magnitude. In this section, we show how the distinct assumptions about the
system boundary and further factors affected the results of the individual studies.

8V.C. Coroama, D. Schien, C. Preist, and L.M. Hilty
Table 1 summarizes the characteristics and the results of the studies pre-
sented. Special emphasis is given to their system boundaries. Our analysis re-
vealed the following factors to be the most important influences on the result:
–The reference year of the study,
–the inclusion of data center devices within the system boundary,
–the inclusion of customer premises equipment and access network, and
–the inclusion of overheads such as cooling and redundancy equipment.
Each influencing factor is discussed in the following subsections.
Table 1. Estimates of the energy demand of Internet transmissions. The columns be-
low “system boundary” show which parts of the Internet and of the end devices (as
introduced in Figure 1) were accounted for by the individual studies. CPE is the cus-
tomer premises equipment. Core stands for the metro and long haul Internet together
– all studies consider both of them. Fibers are the optical transmission lines, and DCs
stands for the equipment in data centers.
Study Method System boundary Data on Energy intensity
CPE Access Core Links DCs [year] [kWh/GB]
[16] Top-down — — X — — 2000 38 – 67
[16], corrected Top-down — — X — — 2000 3.2 – 5.6
[11] Top-down — — X X X 2000 136
[22] Top-down — X X X X 2006 8.8 – 24.3
[23] Top-down — X X X X 2008 7
[24] Top-down X X X — — 2008 0.39
[26] Model — X X X — 2007 0.7 – 2.1
[14] Model X X X X — 2008 >0.179
[12] Model X — X X — 2011 0.006
[29] Model — X X X — 2009 0.02
[31] Meta-Analysis X X X X — 2009 0.057
[32] Case study X X X X — 2009 <0.2
Reference Year. An important part of the large differences can be explained
by the year of reference of the individual studies, ranging from 2000 [16,11]
to 2011 [12]. The ICT sector is characterized by fast innovation cycles, and
the equipment is becoming ever more energy efficient, needing less energy per
amount of data being processed or transmitted. Taking [22]’s estimate that the
energy intensity of the Internet decreases by 30% each year, this technological
progress alone leads to a reduction by a factor of 50 over the period of 11 years
covered by the studies.
System Boundary: Data Centers. The most important determining factor
is the system boundary, in particular, whether or not data center devices (i.e.,

The Energy Intensity of the Internet: Home and Access Networks 9
data storage and server-type devices running in server rooms or data centers)
are viewed as part of the Internet. This decision has a large impact on the result:
As shown in Table 1, for 2008, which is referred to by several studies, the two
studies not including data centers result in energy intensities of 0.39 [24] and
0.179 kWh/GB [14] – factors of 18 and 39 below the 7 kWh/GB of the study
that includes data centers [23], respectively.
The original statistical data from [20] (used by both [11] and [16]) supports
this observation. While the consumption of core Internet devices was estimated
at just 1.25 TWh/a, the consumption of storage devices and servers together
was estimated at 11.7 TWh/a.
System Boundary: CPE and Access Networks. From the equipment in-
ventories in [16] and [11] (presented in Table 1 of each paper), it is clear they
do not consider customer premises equipment or access networks. This stands
to reason: both rely on 2000 inventory data for ICT in US offices, not in homes.
And the Internet had not yet exploded, reaching every home as it does today.
In 2000, ISPs, access networks, and customers’ modems and routers were not
nearly as prevalent or important as they later quickly became. [22] and [23], on
the other hand, explicitly include access networks and exclude CPE. Most of the
other studies account for both, except [27] and [29] which do not include CPE
in their assessment and [12] which does not consider the access network.
Not considering CPE or access networks, however, has a great impact on
the result. When the otherwise dominating data centers are excluded from the
calculation, numerous studies point out that CPE and access networks dominate
the energy intensity of the Internet over core and metro networks. Figure 2, for
example, shows the cumulated power consumption of the case study presented
in [32]. The peripheral parts of the network clearly dominate the overall power
consumption.
Both [14] and [15] also point out that for the peak access rates typical of
2008 and 2010, respectively, access networks dominate the power consumption
of the Internet. Core and metro networks would become the dominating part
only if access rates were to grow extensively and reach peak access rates in
excess of 100 Mbps (for the energy efficiency of 2010 networking technology).
For networking technology that becomes more efficient along the foreseeable
trends, typical peak access rates would have to be over 1 Gbps for the core and
metro networks to surpass the power consumption of the access networks. For
the moment, as typical access rates are much lower, access networks use more
energy than the core of the Internet and their exclusion profoundly changes the
result.
The consumption of customer premises equipment is within the same order
of magnitude as the access network but, as a general tendency, is slightly higher:
according to [14], the access network needs 2.8W of power, while the CPE needs
between 4-10W, depending on the technology used. [31] calculate with 2W per
subscriber for the DSLAM of the access network and with 10 W for the CPE
(5W for the modem and another 5W for the WiFi router). As with the access

10 V.C. Coroama, D. Schien, C. Preist, and L.M. Hilty
network, the decision of whether or not to include the CPE decisively influences
the result.
0
200
400
600
800
1000
1200
1400
1600
1800
2000
0 5000 10000 15000 20000 25000 30000
Cumulated power (excl. PUE) [W]
Distance from Davos [km]
Fig. 2. Cumulated power demand for the videoconference case study [33] along a dis-
tance of 27,000km. The power demand of local network components, albeit relatively
small, is allocated to a relevant extent to the case study’s traffic due to the low average
load of the local components. These local components therefore clearly dominate the
overall power consumption of the transmission. The large core Internet nodes and the
transoceanic “data highways,” while utilizing a relatively large amount of power, typi-
cally have a switching/transporting capacity on the order of hundreds of GB/s, or even
TB/s. Their contribution to the overall demand amounts to less than 1 Watt/Mbps in
the case study. By contrast, the power demand of transcontinental links does contribute
significantly to the overall consumption, due to their relatively low bandwidths.
Overheads: Cooling and Redundancy Equipment. The facilities (rooms,
buildings) hosting ICT networking equipment and datacenters induce a power
overhead due to non ICT-related consumption such as cooling or lighting. The
measure widely used to account for this consumption is the Power Usage Effec-
tiveness (PUE). The PUE is computed as a facility’s total power divided by the
power needed to run the ICT equipment only [34]. As the former includes the
latter, the PUE is larger than, or equal to, 1. The closer the PUE is to 1, the less
power is “wasted” for activities other than information processing. The average
PUE for datacenters nowadays is slightly lower than 2 [35] with a decreasing
tendency; for Internet routers it was around 1.7 in 2009 [36]. Most studies do
consider a PUE between 1.78 [29] and 2 [12, 14, 26, 32] in their calculations. No-

The Energy Intensity of the Internet: Home and Access Networks 11
table exceptions are [16] and [24] that do not account for the PUE, i.e., they
present results one would obtain for a PUE = 1.
Another overhead is induced by redundancy devices. Referring to routers,
[31] notes that “devices are operated with at least dual redundancy in order
to cope with failure.” [14], too, notes that a minimum of two uplinks are used
for redundancy in metro and core networks. Several of the bottom-up studies
(both model-based and case studies) account for the redundancy equipment via
a redundancy factor of 2 [12, 14, 29, 32, 31].
5 Discussion
As the system boundary plays such a decisive influence on the end result, this
section discusses possible best practices in defining the system boundary of the
Internet for energy-related studies. It then shows how the access network and
CPE differ from the more central parts of the network, and concludes with an
assessment of the energy intensity of the former.
5.1 System Boundary
End Devices. [30] argues extensively against the inclusion of end devices (both
user end devices such as personal or laptop computers, as well as servers and
storage in data centers) within the system boundary for the energy intensity of
the Internet. As we have shown above for data centers, including such devices
can dramatically change the results. This, however, is inadequate not only as a
matter of semantics; The concept of the Internet by definition does not include
end devices but only the infrastructure connecting them. There are also practical
arguments against including end devices, since that would yield results ill-suited
and potentially misleading on most questions:
–The consumption in access, metro and core networks is largely independent
of both the end devices and the application generating the traffic. Mean-
ingful averages can thus be defined and estimated. By contrast, end devices
such as desktop computers, smartphones or web servers have very different
power demands. Different applications imply different sets of end devices:
web browsing a server at a data center and a desktop, laptop or tablet com-
puter; peer-to-peer file exchange two computers; and high-end videoconfer-
encing two large LCD screens in combination with codec devices.
–Moreover, even for identical devices at both ends, distinct applications can
induce very different consumption levels per amount of transferred data.
While a peer-to-peer file exchange, for example, can use a bandwidth of
several Mbps, a Skype voice call gets by with a bandwidth of only 60 kbps.
Assuming exclusive usage of the two client devices, the low-bandwidth case
induces a much higher energy consumption per bit at the terminal nodes due
to the low utilization. This could lead to the seriously misleading conclusion
that “the Internet” uses more energy per amount of data for applications

12 V.C. Coroama, D. Schien, C. Preist, and L.M. Hilty
with lower bandwidth demand, in this case for a Skype voice call, than for
a highly demanding file exchange.
Under these circumstances, with different devices and different applications
inducing varying consumption in the end devices, it is unclear how these con-
sumption levels could be aggregated into meaningful averages. It seems more
meaningful always to assess network energy and the energy of end devices sep-
arately, and to add them up when needed – for example, for the assessment of
the energy needs of a specific service [32].
CPE and Access Networks. The consumption of customer premises equip-
ment and access networks, on the other hand, should always be considered.
Unlike end devices, these devices have no stand-alone meaning. They only exist
to connect end devices and thus semantically belong to the Internet.
Moreover, as shown in several studies [14, 15, 32], CPE and access network
dominate the energy consumption of the Internet over metro and core network.
Although this might not be true for services with a very high bandwidth usage
(see next chapter [13]) and might permanently change in the future with in-
creasing access rates [15], it is certainly the rule for the moment. It is advisable
always to include these factors, even if they make an equal contribution in the
cases under scrutiny. Although dropping such factors may simplify comparisons
such as in [12], it will also change the absolute values which, if taken out of
context, may lead to misunderstandings.
Overhead: PUE and Redundancy. The cooling and other types of overheads
included in the PUE, as well as the redundancy equipment, provide support or
safety functions in the Internet. Hence, their consumption should be accounted
for.
As one of their few advantages over bottom-up assessments, top-down studies
inherently include the redundancy equipment, as this is included in the stock
inventories these are based on. Bottom-up analyses must account for them ex-
plicitly – as mentioned above, a factor of 2 for redundancy and a factor of 1.7-2
for the PUE are the values most studies use.
It must be noted, however, that both PUE and redundancy only apply to
access, edge, metro and core equipment. CPE have no redundancy nor are they
cooled. Whether redundancy and overheads are considered thus has a relatively
low influence on end results that include CPE, and there is a risk of overesti-
mation if one includes redundancy and PUE for the CPE as well (as we did in
[32]).
Fibers. Whether the power of the amplifiers along the fiber cables is considered
(as it is in most studies) or not (as in [16, 24]), also has only a marginal impact
on the result. The relatively high power of dozens of kilowatts of transoceanic
fibers gets divided by such a large amount of traffic that the contribution per

The Energy Intensity of the Internet: Home and Access Networks 13
amount of data becomes negligible [32]. The consumption along fibers only be-
comes relevant for fibers with a low load such as the US transcontinental links
in [32]. Such a case leads to a considerable allocation to relatively low amounts
of traffic. Often, however, this is not the case. [14] argues that “the core opti-
cal transport (wavelength division multiplexed links) accounts for only a small
fraction of the total energy consumed by the Internet.” But even when relevant,
the consumption along the fibers was still smaller than the consumption of the
access network [32].
5.2 The Challenge of Defining the Energy Intensity of the Internet
A basic methodological problem with the energy intensity of the Internet is that
for some devices the energy consumption scales with the traffic volume and for
other devices it scales along different dimensions, especially time of usage [31,
30]. The former is true of most networking devices in the metro and core network
[31], while the energy consumption of devices in the access network and CPE (as
well as end devices, which have been excluded from the analysis) usually scales
with the time of usage and is largely traffic-independent [27].
To account for these differences, [31] recommends allocating the energy con-
sumption of a device based on an approach that takes into account the limiting
factor of the device – i.e., the factor that, if increased, would first limit the qual-
ity of service. If the limiting factor is in practice very hard to reach, the approach
allocates energy according to the dimension that, if changed, results in the most
significant change in energy consumption [31]. For the overall network energy
intensity, [30] thus suggests using a combined approach that encompasses inten-
sities defined as both energy per data and energy per time (i.e., power) where
appropriate. We will elaborate on this approach in subsection 5.3.
Existing studies, however, do not differentiate between categories of devices
when defining a metric for the energy intensity. Instead, each study defines the
energy intensity for all devices along the same dimension: Most studies define
the network energy intensity as energy per amount of data [16, 11, 22, 23, 29,32].
With a partial focus on CPE and access networks, [31] defines the energy inten-
sity for those devices as energy per time. Some studies with a focus on access
networks, noting that access networks’ devices are always on and their consump-
tion is thus both traffic- and time-independent, define the energy intensity as
energy per subscriber [26, 27]. Finally, top-down studies can avoid the problem
altogether by computing only the overall energy of the Internet and not relating
it to any other dimension for a measurement of energy intensity (e.g., [24]).
We conclude this chapter by putting forward a formula for the computation
of the energy intensity of the access network and customer premises equipment.
The next chapter in this volume [13] addresses the metro and core networks,
and proposes a formula for the energy intensity of those parts of the Internet.
That energy intensity is defined as energy per data. Summing these two leads to
the first formula for the energy intensity of the Internet that combines an energy
per time component with an energy per data component and thus models reality
more closely than previous work.

14 V.C. Coroama, D. Schien, C. Preist, and L.M. Hilty
5.3 The Energy Intensity of CPE and Access Networks
In this subsection, we develop the formula for the energy intensity of CPE and
access networks. We build on the analysis of [31], which analyzes the energy
consumption of online multimedia services. For the consumption in access net-
works and CPE, that article puts forward the following formula (formula 9 in
the article):
EAN =tsPCP E
NCP E
+PT U
NT U
P U EN et(1)
where
–EAN is the energy consumption in the access network (including CPE in
that article’s terminology) for the consumption of a given service,
–tSthe time of service consumption,
–PCP E and NCP E the power of all CPE taken together and the number of
users connected to them, respectively,
–PT U and NT U the power of the access network devices and the number of
users connected to them, respectively, and
–P U EN et the PUE of the DSLAM, which typically requires cooling.
We start from this formula to define the energy intensity per unit time (i.e.,
power) of customer premises equipment and access networks. For more clarity,
we consider access network and CPE separately and add them back together in
the end. As we are interested in the energy intensity and not the total amount
of energy, the time factor tsdisappears. Additionally, the energy intensity of the
Internet is the average value for one Internet communication and thus always
includes exactly one set of typical customer premises equipment (NCP E = 1) –
what “typical” means in this context will be addressed shortly.
With these observations, and renaming PTU ,NT U , and P U EN et to PAN ,
NAN , and pueAN for more consistency, the intensities of the access network and
the CPE, iAN and iCP E , are:
iAN =PAN
NAN
pueAN (2)
iCP E =PC P E (3)
The trivial formula 3, however, ignores one important aspect that was not
considered in [31] either: the energy used by CPE while idle, i.e., while not
providing any service. This energy has to be somehow distributed among the
services provided during a certain period [37]. We choose to distribute the idle
energy consumption among the services provided over a given period of time
proportionally to the time those services are active.
For the entire cycle over which meaningful averages can be built (i.e., a day,
a week, a year), we define
–tOn the total time in which the equipment is on,

The Energy Intensity of the Internet: Home and Access Networks 15
–tUse the total time in which the CPE is in use, i.e., in which it is used for
data transmission, and
–tIdle the total idle time, when the CPE is on but not used (tI dle =tOn −tU se).
The CPE consumes power for the time tOn but only provides services during
tUse . Distributing the entire power on the services provided during tUse needs
the factor tOn/tU se for extrapolation. With this, formula 3 becomes
iCP E =tOn
tUse
PCP E ,(4)
which, because tOn =tU se +tIdle , can also be written as
iCP E =1 + tI dle
tUse PC P E (5)
The energy intensity of the access network and CPE taken together is
iCP E &AN =1 + tIdle
tUse PC P E +PAN
NAN
pueAN =tOn
tUse
PCP E +PAN
NAN
pueAN (6)
While such a generic formula is interesting from a theoretical point of view,
for added practical relevance, the formula should ideally be parametrized with
meaningful average values. These values should be based on the equipment and
usage patterns considered typical and will change over time.
For the access network, the fraction PAN/NAN represents the energy intensity
per subscriber. For ADSL2+, this was assessed as 3.4W by [14] and 2W by [31].
As [31] is the more recent analysis, we use its value. We further assume a PUE
of 2, following the majority of studies.
For the power of the customer premises equipment (PCP E ), a few older
studies consider only modems [14], while the more recent studies consider both
modems and WiFi routers [24, 31, 37]. We follow [24] who assumes that only a
few users use a modem without a WiFi router and that their number is compa-
rable to those with multiple WiFi routers or WiFi repeaters. This is equivalent
to assuming that 100% of users use both a modem and a WiFi router, either as
two separate devices or integrated into an Integrated Access Device (IAD). Tak-
ing into account the US distribution of IADs versus modems with WiFi routers,
[24] puts forward 7.1W as average CPE consumption for DSL and 9.5W for
cable. The November 2013 Energy Star requirements for small network equip-
ment [38] call for a base power of at most 5.5W for ADSL and 6.1W for cable
IADs, respectively, allowing another 0.8W for fast Ethernet and WiFi, and 0.5W
for the telephony functionality of DSL modems. Considering these numbers as
well, both ADSL and cable IADs are required to be just below 7W. Allowing
for the slightly higher consumption of two separate devices as well as for legacy
equipment, we use PCP E = 8W.
Finding data for the idle and usage times of modems is far more challenging
and is by far the greatest source of uncertainty. A 2011 BBC study [39] found

16 V.C. Coroama, D. Schien, C. Preist, and L.M. Hilty
that set-top boxes (the modems which deliver both cable TV and Internet con-
nectivity) are on for 15.57 hours/day, but it did not address their usage time. A
2007 study found that in Europe, DSL modems are idle for 20 hours/day and in
use for the remaining 4 hours/day [40]. Although this study is older, we use its
assumptions. The study distinguishes between the on and idle state of modems,
and we feel that the on-time of 24 hours/day from [40] better reflects reality
than the 15.57 hours/day from the 2011 study [39]. With these assumptions,
tOn/tU se = 6.
With all these specific values, formula 6 leads to a “currently typical” value
for the energy intensity (per time) of the access network and CPE of
iCP E &AN = 6 ∗8W+ 2W∗2 = 52W(7)
The average value from formula 7 already includes the PUE of the access network
and the idle consumption of the CPE, and can thus be used for quick assess-
ments of energy consumptions in access networks and by consumer premises
equipment. As mentioned above, however, the formula has a relatively large de-
gree of uncertainty, especially due to the uncertainty of the idle time of the
CPE.
6 Conclusion
We have shown that the energy intensity of customer premises equipment and
access networks has to be assessed differently from the intensity of metro and core
networks. We proposed a formula for the intensity of the former, both generically
and parametrized with typical data for 2014. The next chapter [13] complements
this work with a formula for metro and core networks. Taken together, the
two chapters provide an assessment method for the Internet energy intensity
that appropriately uses different allocation approaches for different parts of the
network. Parametrized with typical values for 2014, this method can be used by
practitioners for quick assessments of various Internet-based services.
References
1. Laitner, J.A.: Semiconductors and information technologies: The power of produc-
tivity. Journal of Industrial Ecology 14(5) (2010) 692–695
2. Mattern, F., Staake, T., Weiss, M.: ICT for green – how computers can help us to
conserve energy. In: Proceedings of e-Energy 2010, Passau, Germany, ACM (April
2010) 1–10
3. Hilty, L.M., Ruddy, T.: Sustainable development and ICT interpreted in a natural
science context: the resulting research questions for the social sciences. Information,
Communication & Society 13(1) 7–22
4. GeSI: Smart 2020: Enabling the low carbon economy in the information age. A
report by the climate group on behalf of the global esustainability initiative (gesi)
(2008)

The Energy Intensity of the Internet: Home and Access Networks 17
5. European Commission: Addressing the challenge of energy efficiency through infor-
mation and communication technologies. Communication from the Commission to
the European Parliament, the Council, the European Economic and Social Com-
mittee and the Committee of the Regions 241 final, Brussels (2008)
6. Pamlin, D., Pahlman, S.: Outline for the first global it strategy for co2 reductions.
Technical report, World Wildlife Fund (2008)
7. Erdmann, L., Hilty, L.M.: Scenario analysis: Exploring the macroeconomic impacts
of information and communication technologies on greenhouse gas emissions. Jour-
nal of Industrial Ecology 14(5) (2010) 826–843
8. Hilty, L.M.: Information Technology and Sustainability. Essays on the Relationship
between ICT and Sustainable Development. Books on Demand, Norderstedt (2008)
9. Hilty, L.M., Coroama, V.C., de Eicker, M.O., Ruddy, T., Mller, E.: The role of
ICT in energy consumption and energy efficiency. Technical Report ICT-ENSURE:
European ICT Environmental Sustainability Research; call identifier FP7-ICT-
2007-2 (2009)
10. Coroama, V.C., Hilty, L.M.: Energy consumed vs. energy saved by ICT – a closer
look. In Wohlgemuth, V., Page, B., Voigt, K., eds.: Environmental Informatics
and Industrial Environmental Protection: Concepts, Methods and Tools. (2009)
353–361
11. Koomey, J., Chong, H., Loh, W., Nordman, B., Blazek, M.: Network electricity
use associated with wireless personal digital assistants. Journal of Infrastructure
Systems 10(3) (2004) 131–137
12. Baliga, J., Ayre, R.W., Hinton, K., Tucker, R.S.: Green cloud computing: Balanc-
ing energy in processing, storage, and transport. Proceedings of the IEEE 99(1)
(January 2011) 149–167
13. Schien, D., Coroama, V.C., Hilty, L.M., Preist, C.: The energy intensity of the
internet: Edge and core networks. In Hilty, L.M., Aebischer, B., eds.: ICT Innova-
tions for Sustainability. Springer, Heidelberg, Germany (2014)
14. Baliga, J., Ayre, R., Hinton, K., Sorin, W.V., Tucker, R.S.: Energy consumption
in optical IP networks. Journal of Lightwave Technology 27(13) (2009) 2391–2403
15. Hinton, K., Baliga, J., Feng, M.Z., Ayre, R.W., Tucker, R.S.: Power consumption
and energy efficiency in the Internet. IEEE Network 25(2) (March 2011) 6–12
16. Gupta, M., Singh, S.: Greening of the Internet. In: Proceedings of the 2003 Con-
ference on Applications, Technologies, Architectures, and Protocols for Computer
Communications (SIGCOMM ’03). SIGCOMM ’03, New York, NY, USA, ACM
(2003) 19–26
17. Teehan, P., Kandlikar, M.: Comparing embodied greenhouse gas emissions of mod-
ern computing and electronics products. Environmental Science & Technology
47(9) (2013) 3997–4003
18. Hischier, R., Coroama, V.C., Schien, D., Achachlouei, M.A.: Grey energy and
environmental impacts of ICT hardware. In Hilty, L.M., Aebischer, B., eds.: ICT
Innovations for Sustainability. Springer, Heidelberg, Germany (2014)
19. Chan, C.A., Gygax, A.F., Wong, E., Leckie, C.A., Nirmalathas, A., Kilper, D.C.:
Methodologies for assessing the use-phase power consumption and greenhouse gas
emissions of telecommunications network services. Environmental Science & Tech-
nology 47(1) (2012) 485–492
20. Roth, K.W., Goldstein, F., Kleinman, J.: Energy consumption by office and
telecommunications equipment in commercial buildings – volume i: Energy con-
sumption baseline. Technical report, US Department of Energy, Office of Energy
Efficiency and Renewable Energy, Washington DC (2002)

18 V.C. Coroama, D. Schien, C. Preist, and L.M. Hilty
21. Coffman, K.G., Odlyzko, A.M.: Growth of the Internet. In Kaminow, I., Li, T.,
eds.: Systems and Impairments. Volume IV B of Optical Fiber Telecommunications.
Academic Press (2002) 17–56
22. Taylor, C., Koomey, J.: Estimating energy use and greenhouse gas emissions of
Internet advertising. Technical report, IM C 2(2008)
23. Weber, C.L., Koomey, J.G., Matthews, H.S.: The energy and climate change
implications of different music delivery methods. Journal of Industrial Ecology
14(5) (2010) 754–769
24. Lanzisera, S., Nordman, B., Brown, R.E.: Data network equipment energy use and
savings potential in buildings. Energy Efficiency 5(2) (2012) 149–162
25. Cisco: Hyperconnectivity and the approaching zettabyte era. Technical report,
Cisco, San Jose, CA, USA (2010)
26. Baliga, J., Hinton, K., Tucker, R.S.: Energy consumption of the Internet. In:
Proceedings of the Joint International Conferences on Optical Internet and the
32nd Australian Conference on Optical Fibre Technology, COIN-ACOFT. (2007)
27. Baliga, J., Ayre, R., Sorin, W.V., Hinton, K., Tucker, R.S.: Energy consumption
in access networks. In: Proceedings of the Optical Fiber Communication Confer-
ence and The National Fiber Optic Engineers Conference (OFC/NFOEC), Optical
Society of America (2008)
28. Kilper, D.C., Atkinson, G., Korotky, S.K., Goyal, S., Vetter, P., Suvakovic, D.,
Blume, O.: Power trends in communication networks. IEEE Journal of Selected
Topics in Quantum Electronics 17(2) (2011) 275–284
29. Schien, D., Preist, C., Yearworth, M., Shabajee, P.: Impact of location on the
energy footprint of digital media. In: Proceedings of the 2012 IEEE International
Symposium on Sustainable Systems and Technology (ISSST), IEEE (2012) 1–6
30. Coroama, V.C., Hilty, L.M.: Assessing Internet energy intensity: A review of meth-
ods and results. Environmental Impact Assessment Review 45 (February 2014)
63–68
31. Schien, D., Shabajee, P., Yearworth, M., Preist, C.: Modeling and assessing vari-
ability in energy consumption during the use stage of online multimedia services.
Journal of Industrial Ecology 17(6) (2013) 800–813
32. Coroama, V.C., Hilty, L.M., Heiri, E., Horn, F.M.: The direct energy demand of
Internet data flows. Journal of Industrial Ecology 17(5) (2013) 680–688
33. Coroama, V.C., Hilty, L.M., Birtel, M.: Effects of Internet-based multiple-site con-
ferences on greenhouse gas emissions. Telematics and Informatics 29(4) (November
2012) 362–374
34. Rawson, A., Pfleuger, J., Cader, T.: Green grid data center power efficiency metrics:
PUE and DCIE. Technical Report White Paper #6, The Green Grid (2008)
35. Stansberry, M., Kudritzki, J.: 2012 Data Center industry survey. Technical report,
Uptime Institute (2012)
36. Moth, J., Norris, M.: G ´
EANT3 carbon footprint. In: Terena Networking Confer-
ence. (2010)
37. Schien, D., Shabajee, P., Wood, S.G., Preist, C.: A model for green design of
online news media services. In: Proceedings of the 22nd International Conference
on World Wide Web (WWW’13), Republic and Canton of Geneva, Switzerland,
International World Wide Web Conferences Steering Committee (2013) 1111–1122
38. EPA: Energy star program requirements for small network equipment. Technical
report, US Environmental Protection Agency (November 2013)
39. Chandaria, J., Hunter, J., Williams, A.: A comparison of the carbon footprint of
digital terrestrial television with video-on-demand. Technical Report WHP 189,
British Broadcasting Corporation, London, UK (March 2011)

The Energy Intensity of the Internet: Home and Access Networks 19
40. Nissen, N.F.: EuP preparatory study lot 6: Standby and off-mode losses – task
3: Consumer behaviour and local infrastructure. Technical Report TREN/D1/40,
Fraunhofer Institute for Reliability and Microintegration, IZM, Berlin, Berlin, Ger-
many (October 2007)