|
|
| Fig. 1: Server hardware at a data center. (Source: Wikimedia Commons) |
Over the last few decades, the internet has become an overwhelmingly important facet of daily life. This has been enabled by the creation of a large number of data centers (buildings hosting large numbers of computers and servers that store the data and perform the computations necessary for the modern internet to function. (See Fig. 1.) As global internet traffic continues to increase rapidly, the energy consumption of these data centers (and the internet as a whole) becomes increasingly important.
Unfortunately, the complicated nature of the internet (and the fact that it is dominated by private corporate entities) means that there is a scarcity of publicly available, high quality measurements of internet power consumption and efficiency. This problem is particularly difficult for trying to understand user-end devices (things like personal computers and smartphones that are connected to the internet) and networks (both mobile and fixed-line networks that actually transmit data). Because of this, we will primarily focus on the energy usage of data centers.
In the interest of completeness, we will begin by discussing internet-connected devices and data transmission networks. We will then move on to analyze the fact that data center energy usage remained relatively steady during the period from 2010 through 2018. Finally, we will consider the future of data center energy requirements.
Before we move on to discussing data centers, we will briefly summarize some results about the power consumption of user devices and networks. The typical lifespan of an internet-connected user device, such as a personal computer or a smartphone, is only a few years. Furthermore, many of these devices (like phones, laptops, and tablets) are relatively small. The combination of these factors results in high energy efficiencies for most user devices; in fact, some studies estimate that the total energy used by these devices throughout their lifetimes is substantially less than the amount of energy required to manufacture them. [1] These battery-powered devices are anticipated to be a less important contributor to the energy consumption of the internet than smart appliances (which are increasingly popular and use energy very inefficiently in their standby modes), network connections, and data centers. [1]
While network traffic has increased significantly over the last fifteen years or so, it has also become more efficient. As mobile networks have upgraded from 2G to 4G and even 5G, their efficiency has improved by nearly two orders of magnitude. [1] Similarly, some analyses estimate that the efficiency of fixed network transmission lines has doubled roughly every two years. [2] Unfortunately, there are fewer analyses and forecasts available about the changes in network power requirements than there are about data centers. Consequently, we will turn our attention to the energy usage of data centers for the remainder of this work.
While the information load on data centers increased exponentially from the mid 2000s through 2018, the energy consumption of these data centers increased only modestly. Despite a 550% increase in computational load over this time period, the energy consumption increased by only 6%. [3] In absolute terms, data centers used roughly 194 TWh of energy in 2010 and roughly 205 TWh of energy in 2018. [3] In both years, this was approximately 1% of global electricity consumption. [3]
This remarkable stability in energy usage despite radical increases in computational load was due to efficiency gains in the hardware, software, and infrastructure (cooling, etc.) of these data centers. For example: from 2010 to 2018, improved processor efficiencies caused the energy per computation of a typical server to decrease by a factor of four, and improved efficiencies in storage drives caused the power per terabyte of stored information to decrease by a factor of nine. [3] A software innovation called virtualization (which essentially allows one physical server to function as several different servers for computations) caused the number of calculations a typical server could host to increase by a factor of 5 from 2010 to 2018. [3] Finally, the power usage efficiency (PUE), which is the total power consumed by a data center divided by the power used by its IT devices, has dropped significantly since 2010. In 2010, a typical data center had a PUE of 1.9, meaning that almost half of the energy used by the data center went to infrastructure (cooling, lighting, backup generators, and so on). [4] In 2018, a typical data center was estimated to have a PUE of 1.55 or so. [4]
These improvements have largely been achieved by a shift in the typical size of a data center. Since 2010, much of the information volume handled by U.S. data centers has been moved to hyperscale data centers. These are exceptionally large data centers that handle web traffic for extremely large users (like Amazon, Google, and others). The profit incentives of these data centers, combined with their specialization and economies of scale, have encouraged many of the improvements in energy efficiency described above. While little has been published about the measured energy efficiencies of data centers since 2018, there have been semi-empirical analyses that combine physical models of data centers with previous energy measurements to estimate the PUE of modern data centers. One such analysis published in 2022 estimates that hyperscale data centers have a median PUE of 1.12- 1.25, midsize data centers have a median PUE of 1.39-1.98, and that small data centers have a median PUE of 1.71-2.22. [5] This supports the argument that a transition to hyperscale centers has been an important driver of improved energy efficiency. Furthermore, these values lend credence to industry claims that (under optimal conditions) some of the most efficient modern data centers can achieve PUEs of less than 1.1. [5]
However, we should also note that energy usage of data centers is difficult to estimate. Not only are such analyses published relatively infrequently, different authors have dramatically different estimates of energy usage. For example, one recent study cited the global energy usage of data centers in 2022 as 416 TWh. [6] This number is significantly higher than Masanet et. al's 2018 estimate of 205 TWh. [3] It is unclear how much of this difference is actual growth in energy usage and how much is due to differences in methodologies. Since this energy consumption cannot be measured directly by the public, estimates rely upon bottom-up modeling (which considers the small- scale structure of data centers, like servers and hardware, to create a model of a data center) or top-down approaches (which do not consider the specific architecture of a data center and instead rely upon summary statistics). [7] It has been found that bottom-up approaches are typically more accurate. [7] However, that accuracy relies upon more detailed knowledge of the data centers. Clearly, the amount of detail needed for a precise bottom-up model of the global data center network is unreasonably large. Additionally, we note that none of the authors whose total energy estimates cited above have offered error bars on their estimates. Consequently, we should suspect that both modeling methods come with large uncertainties and that the true value of energy usage is likely somewhere in between these estimates.
Given these considerations, the exact values (particularly estimates of total energy usage) listed above should be taken as very rough estimates. However, we note that the high-end 2023 estimate of 416 TWh differs only by a factor of roughly two from the low-end 2010 estimate of 194 TWh. [3,6] Presumably, much of this difference is due to differences in the models of data centers. Thus we conclude that the increase in energy consumption of data centers from 2010 to 2018 was significantly less than double. Then, even if the claim that computational load increased by 550% between 2010 and 2018 is quantitatively wrong, we can be confident that the actual value of the increase was significantly more than a factor of two. As a result, we can reasonably conclude that whatever the exact value of the increase in energy consumption, it grew far less than computational load increased. This must have been due to significant efficiency gains. Even though the calibration of the bottom-up models is probably incorrect, their self-consistency leads us to believe that their explanation of the efficiency gains is likely a good approximation to the truth.
It seems likely that societal demand for data storage and cloud computation will continue to increase rapidly. Indeed, the rising popularity of cloud-based artificial intelligence models could cause the growth of information traffic handled by data centers to accelerate further. In order to keep these demands from causing a dramatic increase in the power consumption of the internet, we must continue to improve the efficiencies of network connections and especially the efficiencies of data centers. Historic trends indicate that consolidating information traffic to hyperscale centers, combined with further advances in hardware, software, and infrastructure, are an effective way to deal with this problem. Models of the energy requirements of U.S. data centers support this idea, as well as suggesting that significant efficiency gains can be made by removing inactive servers. [8]
Dramatic efficiency improvements in user devices, data transmission networks, and data centers have kept the energy usage of the internet at a relatively steady value even as the world has become increasingly digitized. For this to continue, the information industry must continue to pursue efficiency gains in a variety of areas. Data on these improvements (particularly after 2018) is rather sparse; furthermore, different models that incorporate these data have significantly different estimates of total energy consumption. More frequent and more detailed reporting might allow for a consensus understanding of this energy usage to be reached (and error bars of these estimates to be given). In turn, this understanding would allow for more informed public and corporate policies to encourage the necessary efficiency gains. Given the increase in internet traffic, as well as the spreading popularity of smart appliances and artificial intelligence, these policies will become an increasingly important part of any societal effort to manage global energy usage. It is therefore of great societal importance to continue these efficiency gains and to collect better data in order to understand these improvements.
© Nicolas Ferree. The author warrants that the work is the author's own and that Stanford University provided no input other than typesetting and referencing guidelines. The author grants permission to copy, distribute and display this work in unaltered form, with attribution to the author, for noncommercial purposes only. All other rights, including commercial rights, are reserved to the author.
[1] "Digitalization and Energy," International Energy Agency, 2017.
[2] J. Aslan et al., "Electricity Intensity of Internet Data Transmission: Untangling the Estimates," J. Ind. Ecol. 22, 785 (2017).
[3] E. Masanet et al., "Recalibrating Global Data Center Energy-Use Estimates," Science 367, 984 (2020).
[4] A. Shehabi et al., "Data Center Growth in the United States: Decoupling the Demand for Services From Electricity Use," Environ. Res. Lett. 13, 124030 (2018).
[5] N. Lei and E. Masanet, "Climate- and Technology-Specific PUE and WUE Estimates for U.S. Data Centers Using a Hybrid Statistical and Thermodynamics-Based Approach," Resour. Conserv. Recycl. 182, 106323 (2022).
[6] S. Long et al., "A Review of Energy Efficiency Evaluation Technologies in Cloud Data Centers," Energy Build. 260, 111848 (2022).
[7] R. Bertran et al., "Counter-Based Power Modeling Methods: Top-Down vs. Bottom-Up," Comput. J. 56, 198 (2013).
[8] A. Shehabi et al., "United States Data Center Usage Report," Lawrence Berkeley National Laboratory, LBNL-1005775, June 2016.
