So the quick dive into shedding light on the dark "slide" (eek) of the webpage in pt I, becomes a full on jump down a rabbit hole in pt 11, to look at the footprint of all the magic that brings the page to the screen.
Again excuse the title, paraphrased from Batman's line in Lego Movie and less relevant here, but it connects the posts and i'm all out of puns for now.
There's a whole bunch of analyses of the energy demand and carbon emissions associated with web, on the web (it eats itself!). From sender request to server response, from the movement of data, to the download and rendering of content, and everything in between; it will have been covered. Either in full life-cycle analysis (LCA) articles [1], or search engine optimisation (SEO) posts [2] the numbers will have been crunched.
If you want rigour look elsewhere, but hopefully the numbers played with here help to illustrate the kind of data that can be drawn from the web, and how it can be used to derive useful insights into how clean our feet really are. Key takeaways...
Top-down network energy intensity ~ 0.16 kWh/GB.
Bottom-up, ditto ~ 0.15 kWh/GB and user device ~ 0.09 kWh/GB.
Energy of a "click n read" of an average 2.3 MB page and 54 s read ~ 0.84 Wh.
Emissions per click ~ 0.34 gCO2.
So a London to New York flight is nearly 1 M clicks, a km drive nearly 500 views.
The power of the click
Boundaries blur when it comes to energy intensity of the web, with numbers varying over 4 orders in magnitude, 0.004 to 136 kWh per GB of data moved through the network, depending on the boundaries on the analysis [3].
First order estimates of top down and bottom up can be made and compared for consistency. IEA assessed data centre and transmission energy to be around 300 TWh (+/- 50) each in 2022, with 4.4 ZB of data pushed through the network [4].
Shockingly, crypto mining adds another 125 TWh (+/- 25) to the load, giving an overall intensity of 725(kWh)/4.4(ZB) ~ 0.16 kWh/GB (+/- 0.02). Against best practice and con"science", uncertainties are dropped from here only for simplicity.
Bottom up is more complex and requires assessment of device power, time in use, data speeds and the number of "hops" that data takes through all components on the network. No easy task, but fun to try at a guestimate!
In the data centre, bytes pass through a server and processor each rated around 600 W, supported by drives and PSUs of the same ratings [5, 6]. Cooling the units demands an equivalent to the sum of these powers again, approx. 2.4 kW. Backup storage, lighting, networking account for near 20% of total demand, giving a total of 6 kW.
Connection through the centre is via ethernet at speeds of 1-100 Gbs. Assuming a low end of 1 Gbs, the energy intensity is then 6(kW) x 8(bits/byte) / 1(Gbs) ~ 48 kWs/GB, or x 1/3600(s/h) ~ 0.013 kWh/GB per "centre".
Number of hops via these centres depends on locations of "sender" and "server", but not so much on distance, and can range 2-10 or more (check with >traceroute or >tracert or on the web, e.g. https://traceroute-online.com/).
Let's parallel the theory of "six degrees of separation" [7], then for 6 hops the contibution from data centres to the overall energy rounds to 0.08 kWh/GB, matching the top down number, but clearly subject to uncertainties in hop number and server coponent powers.
The message needs repeating
Transmission is a little more complicated, but fun to make a calculated guess. Bits and bytes travel the net as light (or electrons) through optical fibres (or copper cables), which can be lost over the large distances that fibre connections make.
To avoid losses, the signal is boosted by repeaters placed every 50 km or so along the fibre cable, powered by a +7.5 kV end to -7.5 kV end supply at 2A, or 30 kW [9]. Check refs [10] for "deep" insights on the subsea network, vulnerabilties. Assuming a bit rate through the fibre optic of 1 Gbs then the typical energy intensity is 30 kWs/Gb or 0.07 kWh/GB.
Pushing this further, a handle on the typical distance that data travels on the net is derived from the speed of light (300000/n ~ 200000 km/s in glass fibre, n ~ 1.5) and "average" ping time to send data across the net.
You can go pinging sites across the world to be build the stats on round trip times (RTTs), or simply grab data from a site that does this for you [11].
A typical RTT is around 20 ms, but includes a latency of 1 ms(ish) processing/routing time in the data centre at each hop. So a "through cable" time of order 14 ms for our six hop example. This gives a one way distance of travel at 200K x 0.007 ~ 1400 km, and unit energy of 0.04 W/GB/km [cf. 12].
To the point, putting our bottom up approach to data centre (0.08 kWh/GB) and transmission (0.07 kWh/GB) energy intensities together gives 0.15 kWh/GB, a value close to the top down approach of 0.16 kWh/GB.
No claims on rigour here, just ball-park figures, born out of a curiosity in the data that feeds the calculation and the components that make up the net user's footprint.
Not bad given the web is more complex than these simple calculations portray. For example, if data transmission is limited only to copper lines, where bit rates drop to 300 Mbs, then energy intensity increases to 0.06/0.3 ~ 0.2 kWh/GB.
It all adds up
Back to the click and more boundaries. For a tap on the mouse/pad/screen that retrieves a typical 2.3 MB webpage [13], the energy cost of a click comes in around 0.16(kWh/GB) x 0.0023(GB) ~ 0.4 Wh per page in data centre and transmission.
The IEA's estimated 5.3 B internet users averaging 6.7 h a day on the web, of which 2.7 h is spent streaming video [4, 14, 15]. This leaves 4 h of "clicks n reads" and with an average read time of 54 s per page [16], the total network energy cost is 5.3B x 0.4(Wh) x 4(h/d) x 3600(s/h) x 365(d/yr) / 54(s) ~ 200 TWh each year.
This makes up 200/600 ~ 33% of the total energy consumed in data centres and transmission (excluding crypto) [4], a value consistent with video streaming now taking a 66% share of web traffic [17].
Images make up nearly 60% of the page weight [13], so scope for reducing footprint through image compression, reduced resolution, a single composite download and cropping script for the different images on the page. Guilty as charged here, web optimisation's for another day!
In any case, if we expand the boundary to include the energy of the end user device for the average read time on typical 50W laptop, 50(W) x 54(s) / 3600(s/h) ~ 0.75 Wh, then the total energy demand is around 1.2 Wh per page (click + read).
Pic downloads over the net (data centre + transmission) are then only 60% x 0.4 / 1.2 ~ 20% of the overall energy. OK, every little counts, guilty of gifs and could do better. Apparently, much better according to websitecarbon.com (eek)!
User device ratings play a major role in the overall energy demand, with mobile (1 W) and desktop (250 W) ranging the energy per read 0.015-3.75 Wh.
In fact, increasing mobile use reduces the overall user device footprint into the future, grabbing 60% of internet traffic in 2023, with laptops/tablets taking 40% [18]. This gives a weighted end user energy of 0.6 x 0.015 + 0.4 x 0.75 ~ 0.3 Wh per page.
The calculation excludes the embodied energy of production servers, cooling units, cables, repeaters, and all devices that make the net... work (aha), as well as the energy required to transport and deploy them!
But hey, we have to stop somewhere and estimates place production at another 250 TWh per year with intensity 250(Twh)/4.4(ZB) ~ 0.06 kWh/GB, or x 0.0023(GB/page) ~ 0.14 Wh per page [19].
Summing the energies from data centre, transmission, user device and production gives 0.2 + 0.2 + 0.3 + 0.14 ~ 0.84 Wh per page. Then for the average page size (2.3 MB) the energy intensity is around 0.4 kWh/GB, a value that sits neatly in the middle of the four order of magnitude range of estimates we started with [3]!
And finally... using our global average emissions of 400 gCO2/kWh [see pt I], we can now estimate the footprint for each contribution to the click! Date centre and transmission rolls in at 0.08 gCO2 each, end user device at 0.12 gCO2 per click and read, and embodied emissions around 0.06 gCO2, with a total coming in around 0.34 gCO2 per page.
For context, a London to New York flight produces 300 kgCO2 per passenger [20], equivalent to 300K (gCO2) / 0.34 (g/page) ~ 0.9M pages or nearly 20 yrs of reading at the US average of 130 webpages a day [21].
Scaling to global users gives emissions of, 5.3B x 0.34(gCO2/page) x 130(pages/d) ~ 0.23 MtCO2 a day, or x 1000/300/200 ~ 3900 LHR to JFK flights a day, each with a full 200 passenger load.
Alternatively a km drive in a gas guzzling car is equivalent to 164(gCO2/km) / 0.34(gCO2/page) ~ 482 pages/km, nearly 500 page views!
We can estimate click numbers from the 4 hours spent online (not streaming video), which gives 4(h) x 3600(s/h) / 54(s/page) ~ 267 per day (per person).
Evidently, this returns the top down energy of data centre and transisson as 0.4(Wh/page) x 267(pages/d) x 365(d/yr) x 5.3B ~ 200 TWh/yr, which at 34% of network traffic devoted to data other than streaming or crypto, gives a total network energy of ~ 600 TWh per year (cf. above)
Can we slice this a different way? "A minute on the internet" from Domo is a popular and tangible measure of web activity [22, 23]. Summing non-video content, including emails, social media and searches gives nearly 300 M clicks per minute, or x 60(min/h) x 24(h/d) ~ 432 B a day.
The number of users online at any moment over the course of the day, given the average 6.7(h) spent online [14] is, x 5.3B(users) / 24(h) ~ 1.5 B, such that clicks per user, come in at 432B(clicks/d) / 1.5B(users) ~ 288 clicks a day.
Makes sense too when we look at the typical monthly energy bill, 0.84(Wh/click) x 288(clicks/d) x 31(d/mth) ~ 7.5 kWh/mth, a value well in range of any online energy calculator [24], when the mix of user devices and boundaries are considered.
So the numbers appear to add up rather well and certainly match stats from my history tracker (though i've been a bit click happy here searching data).
Take home is that everything we interact will have a footprint on the world, the climate and nature. That print can be large when boundaries are broadened and the impact of the individual is scaled to the masses.
So, whether it's a click on the web, or the coffee cup you're about to slam down in relief that this post is over and done, stop and ponder just how big and filthy your feet really are!
https://www.wholegraindigital.com/blog/website-energy-consumption/
https://www.iea.org/energy-system/buildings/data-centres-and-data-transmission-networks
https://www.deltapowersolutions.com/en/mcis/tools-estimated-power-rating-security-appliance.php
https://www.clarke-energy.com/applications/data-centre-chp-trigeneration/
https://www.racksolutions.co.uk/news/blog/how-many-servers-does-a-data-center-have/
https://hackaday.com/2023/08/08/under-the-sea-optical-repeaters-for-submarine-cables/
https://www.cnet.com/home/internet/features/the-secret-life-of-the-500-cables-that-run-the-internet/
https://thundersaidenergy.com/downloads/energy-intensity-of-fiber-optic-cables/
https://www.statista.com/statistics/1380282/daily-time-spent-online-global/
https://www.sandvine.com/inthenews/netflix-eats-up-15-of-global-downstream-traffic
https://www.mobiloud.com/blog/what-percentage-of-internet-traffic-is-mobile
https://www.akcp.com/blog/the-real-amount-of-energy-a-data-center-use/; https://www.mdpi.com/2078-1547/6/1/117#abstract
https://www.icao.int/environmental-protection/Carbonoffset/Pages/default.aspx
https://www.demandsage.com/website-statistics/#:~:text=However%2C%20on%20average%2C%20internet%20users,different%20web%20pages%20per%20day.
https://www.statista.com/statistics/195140/new-user-generated-content-uploaded-by-users-per-minute/
*Links accessed at 05/02/2024