How Much Is Netflix Really Contributing to Climate Change?

The conversation got underway slowly. It began last July, when the Paris-based think tank The Shift Project published a report [PDF] advocating a kind of social state of mind it called “digital sobriety.” The report alleged that electricity expended in the act of delivering streaming video during 2018 (including data center power) was equal to about 96 percent of electricity used in all of Spain in 2011.

Then, in October, the online issues blog Big Think brought more publicity to the report.  The blog quoted a member of The Shift Project as saying that Netflix’s streaming of a half-hour-long video to one viewer resulted in the same amount of CO₂ emissions as driving an ordinary gasoline-powered automobile four miles.

The debate picked up steam this January, when Stanford University consulting professor Jonathan Koomey weighed in on Twitter. Citing a corollary of widely-held industry observation that has come to be known as Koomey’s Law, Dr. Koomey called foul:

Big Think has since heavily revised the blog page that started all this, removing many of the claims and dramatically shrinking the size (as well as the claims) of the article.

Base Levels

The fact that there is a debate on this topic, at this general a level, at this stage of the global climate change crisis, speaks to how little we all know about the evolving, perhaps metamorphosing, topics of data center power and carbon emissions. On one side of the argument is the genuine disbelief, backed up by data (which we’ll share momentarily) in the theory that every viewing of a TV sitcom triggers the emission of as much carbon as a drive to the hospital. Perhaps no industry has done more to consistently re-engineer its core architecture for measurably reduced power consumption than the data center industry.

Related:Here’s How Much Energy All US Data Centers Consume

Yet there’s some validity to the report’s central concern. When television was a broadcast industry, transmitters may have had tremendous power boasted by their stations in “thousands of watts!” Yet they served tens of thousands of viewers simultaneously. Certainly, video streaming, being a data center-driven task, must consume substantially more power than wireless broadcast transmission in serving the same number of viewers.

Are we being too careless in not knowing just how many kilograms or pounds of CO₂ we’re talking about off the top of our heads?

“To do these kinds of calculations, you have to have a certain base level of technical knowledge,” Koomey said in an interview with Data Center Knowledge. “And you have to have experience with the kinds of problems that can come into these calculations.”

Related:Want Microsoft as Your Customer? Clean Up Your Carbon

fulton streaming chart 1

In 2017, Koomey co-authored a study, published in the Journal of Industrial Ecology, investigating the energy consumption of all classes of data being delivered via internet by data center servers. That study produced the latest rendition at the time of the phenomenon of Koomey’s Law: The electricity intensity of internet transmissions, measured in kilowatt-hours per gigabyte (kWh/GB), were continuing to drop at a logarithmic rate, with no end in sight. Using observed data, the team calculated that in 2015 data centers consumed 0.06 kWh/GB, compared to about 6.8 kWh/GB in 2000. At the rate of decline measured then, data centers would have stopped using electricity altogether before the end of 2019.

That being impossible, Koomey told us, it’s fair to assume that data centers today, at the start of 2020, consume as little as 0.015 kWh/GB in the transmission of data through the internet to a user. This would be an average for all classes of data-driven applications, including but not limited to streaming video.

By Netflix’s own public estimates, a one-hour-long 4K Ultra-HD video consumes 7GB of transmitted data on average. Using Koomey’s current-consumption estimate, that half-hour stream should consume 0.0525kWh of electricity.

As DCK readers know full well, the world’s data centers are not entirely fueled by coal-burning generators. In Koomey’s experience researching all forms of energy generation, a common rule of thumb is that every 1kWh of energy generated by coal yields 1 kg of CO₂. But according to more direct measurements, in a hypothetical all-coal world, every kilowatt-hour would yield 52.5 grams of CO₂ emissions.

For its report on online video, The Shift Project chose to use the energy conversion factor of 0.519 kg CO₂ e (carbon dioxide equivalent) per 1kWh. Up until recently, that figure had been used by the UK Government’s Building Research Establishment as a 2014 estimate of energy emissions on account of electricity consumed by residential dwellings, including bungalows and huts.

But citing the dramatic new energy strategies applied to the UK’s energy grid since that time, British regulators in January 2019 adopted a downwardly revised figure of 0.233 kg CO₂ e per 1kWh. It’s important to note, however, that this figure was never intended to be applied to data centers. Furthermore, it’s not a worldwide estimate, as The Shift Project claimed, but a rule-of-thumb number applied mostly to British households.

Soap Opera

Is there a conversion factor that applies not just to the specific case of data center power but the narrower use case of serving streaming video? For the answer, we found Andrew Sauber, an engineer with cloud platform provider Linode. His previous experience includes responsibility for servers with the express purpose of streaming video — specifically for what was until 2018 the world’s principal online destination for South Korean soap operas.

In Sauber’s experience, each physical CDN server maintained a 40 Gbps link to the internet. Each connection to that server consumed 7,500 kbps of bandwidth. The power draw for each server, he reported, was 321W. Using that math, at full saturation (for which each operator was instructed to strive) each server would be providing 5,333 simultaneous streams.

By Sauber’s estimate, the server would consume 160.5 watt-hours to serve a half-hour video to more than 5,000 viewers simultaneously. Using that math, a half-hour soap opera streamed from his server to each viewer would consume 0.0301 watt-hours.

The Shift Project’s projection attempted to include all carbon emissions for energy generated through the entire streaming video delivery lifecycle, which would encompass internet transmission power, including routers and switches for all intermediate hops, as well as the device the viewer used. An ordinary 49-inch diagonal smart TV (one that connects to Netflix directly, without a dongle or attached PC) consumes 145W. As for internet transmission power, Koomey provided an estimate based on studies in which he directly participated: about 0.07kWh for the half-hour period.

Adding it all up, streaming half an hour of a Korean soap opera takes 0.1426 kWh of electricity. Sauber computed that an average light-duty automobile traveling four miles should consume some 6.28kWh. According to Natural Resources Canada, an internal combustion engine burning 1 gallon of gasoline produces about 8.7 kg CO₂ e. A nearly four-mile drive in Sauber’s light-duty car would produce 1.52 kg of carbon emissions.

Using the corrected version of the metric Shift Project chose, that half-hour soap opera would produce about 0.0332 kg CO₂ e. That puts The Shift Project’s estimate off by a factor of about 45. Put another way, every component in that soap opera streaming process would collectively pump the equivalent amount of carbons as that light-duty vehicle traveling about 461 feet. That’s about as long as the City Hall Building in Los Angeles (the one on the “Dragnet” police badge) is tall.

Los Angeles City Hall (Image: Michael Fromholtz)

Realizations

Even these results have probably become dated by current trends. Unlike Sauber’s old digs, Netflix uses a revolutionary microservices-driven server architecture its engineers invented. The old CDN was designed around the first generation of virtual machines, probably during the 2000s, when containerization was a class of advice given to co-workers who shared the same refrigerator in the break room. Today, Sauber’s Linode environment is highly containerized. Appliances that once handled the task of load balancing have been replaced by software-based microservices such as NGINX and HAProxy that distribute workloads much more efficiently, and at a very granular level.

“The reason that containers are so much more efficient than virtual machines is that they don't need to pretend to be a collection of physical hardware,” Sauber wrote in a note to DCK. “They simply use a variety of Linux kernel features (cgroups, network namespaces) that isolate the filesystem and network of a particular workload.”

Sauber personally contributes to these increases in efficiency as a contributor to Kubernetes, through his membership in the CNCF’s Cluster Lifecycle SIG. There, he’s personally witnessed efficiency gains in orders of magnitude, the precise measurements of which have yet to be fathomed.

So, the actual amount of carbon emissions resulting from operations of Netflix and Amazon Web Services (on top of which much of Netflix runs) is probably even lower than our estimate above.

How could The Shift Project overestimate by so much? Koomey noted that the French think tank often cited the work of Huawei Technologies researcher Dr. Anders Andrae, whose 2015 estimate of global electricity usage [PDF] speculated that by 2030 communications tech could be responsible for consuming more than half the total electricity generated in the world.

As Koomey told us, Andrae’s methodology is based on a forecast of projected increases in the world’s consumption of data. That forecast was based on multi-fold increases estimated at the start of the last decade. Andrae then, Koomey noted, applied blanket electricity-consumption figures against data consumption, and then increased his energy projections by the same factor as his data projections – without taking account of the efficiencies and innovations attained in both data distribution and energy consumption in recent years. In other words, the Andrae study appears to assume that a kilowatt-hour five or ten years from now will have the same efficiencies as a kilowatt-hour today.

What have we learned from this exercise? It should not take an under-researched overestimate of ITC energy usage or carbon emissions for people in the ITC industry to take serious note of how much power data centers, and the technology surrounding them, consume. Granted, servers are not powered by V8 engines or Franklin stoves, nor are they spitting out carbons as though they were.

fulton streaming chart 3 corrected.png

But if we are to take activist Greta Thunberg’s admonitions seriously, then we need to have a comprehension of the systems we’re using and the energy they utilize at least as thorough as Andrew Sauber’s. If we let folks on the talk-show circuit take over this topic, they will pollute the conversation as rapidly as greenhouse gases have polluted Earth’s atmosphere. Then, when we need the support – and eventually the good faith – of people in government, we’ll find them already hijacked by extremists.

Driving a gasoline-burning car even 460 feet still contributes to the greatest threat this planet faces. Now that we know it’s 460 feet, we have a better chance of making it 230, and from there, 115.