IEEE Circuits and Systems Magazine - Q1 2023 - 66

where tr is the duration of the video corresponding to
J
refers to the energy consumption needed to decode one
second of video content, which is given in the unit svideo
for the r-th request on the σ-th server. Note that in the
case of real-time decoding, which is not required in DCs,
this metric is the same as the decoding power consumption
in Watts.
Considering the encoding energy consumption, different
models are presented in [66], [67], and [68], which
show that it depends on the same parameters (e.g., resolution,
frame rate) as the decoding energy consumption.
Further works target the optimization of the encoder energy
consumption [69], [70], [71], [72], [73], where additional
parameters influencing the energy consumption
have been derived. Next to the parameters mentioned
for decoding, these are mainly the encoder configurations
that have a significant influence on the encoding
complexity. Examples are the bit stream configuration
such as all-intra coding or random-access coding [74]
and the choice of the target bitrate or the target visual
quality using a quantization parameter (QP) [75]. Additionally,
many encoders (e.g., x264 [76] or x265 [77])
can be tuned, where so-called presets and other encoding
parameters can be chosen to control the encoding
speed [18] or the target compression performance.
Note that these parameters can also affect the power
consumption of recording videos on end-user devices
(5), as the encoding of videos is usually done simultaneously
with capturing. To be consistent with the model
for the decoding energy, we model the encoding energy
consumption for a single video as
ep tσσ=⋅ (13)
sr vsVP transenc
VP transenc
,, ,, ,, ,, ,, ,,
rv r ,
where ps,VP,σ,trans,r,enc,v denotes the encoding energy per
video second of the v-th video to be encoded. In [78], it
is shown that this kind of energy model is sufficiently
accurate (estimation error below 8% on average).
After transcoding, the videos need to be stored on
the servers for permanent availability. The corresponding
energy consumption was discussed and modeled in
[17] as follows:
Ee Bss
vVs,, ,
VP storeVPstore
,, ,, ,,
σσ
where es,VP,σ,store in bityear⋅
J
=⋅ ∑
∈ VP store
σ
is the energy consumed for
storing a single bit on the σ-th server for one year. v indicates
one video of the set of videos Vs,VP,σ,store stored on
the σ-th server.
Table 3 lists energy consumption values taken from
the literature. For the PUE η, we choose a constant
66
IEEE CIRCUITS AND SYSTEMS MAGAZINE
v,
(14)
the r-th transcoding request. Here, ps,VP,σ,trans,r,dec in svideo
Table 3.
Energy parameters for video providers. The encoding
power consumption and the decoding power
consumption are listed for an HD video at 30 fps, and a
bitrate of 2.
Mbit
s
Parameter
η
Es,VP,σ,0
ps,VP,σ,trans,r,enc
ps,VP,σ,trans,v,dec
es,VP,σ,store
es,VP,σ,send = es,VP,σ,Rx
Value
1.08
127 kWh − 5.57 MWh
200 mJ 90 s
kJ
−
svideovideo
719 mJ 24.45 s
−
0.59 Wh
MByte
0.624
mWh
MByte
J
svideovideo
per year
Source
[29]
[32]
[73], [79]
[47], [64]
[17]
[17]
value for all servers which was reported for a social
network in [29]. For the encoding energy consumption,
we provide a range of potential values because
it highly depends on the used technology and the
implementation of the encoder. It is worth mentioning
that the low value for encoding one hour of video (200
mWh) is more than five orders of magnitude smaller
than the high value (90 kWh). The reason for this
extreme difference is that for the low end, a highly
optimized hardware encoder chip customized for low
power applications was used, whereas for the high
end, an extremely complex software implementation
targeting a maximum compression performance was
employed. A similar range is given for the decoding
power consumption, where the lower value is reported
for hardware decoding and the higher value
for HEVC software decoding, both on an evaluation
platform.
Finally, we note that similar to end-user devices, also
the energy consumption values in DCs differ significantly
depending on the deployed hardware components
(CPUs, GPUs, FPGAs) and the video properties. For more
accurate models, these factors need to be taken into account,
especially as energy values can differ by several
orders of magnitude (e.g., those for video encoding as
mentioned above). However, in this article, we focus on
the reported range of values instead of a more accurate
modeling approach to obtain upper and lower bounds
on the total energy consumption. A more accurate modeling
approach and its implications can be considered
in future work.
D. Transmission Networks
For the yearly energy consumption related to transmission,
one approach is to separately consider all network
FIRST QUARTER 2023
IEEE Circuits and Systems Magazine - Q1 2023

Table of Contents for the Digital Edition of IEEE Circuits and Systems Magazine - Q1 2023
