Sustainable Networking (Context)

Welcome everyone for the last block of Adv-Net. My name is Romain. I'm a postdoc working with Laurent for a bit more than four years now, and more specifically on sustainable networking for a bit more than two years.

We will spend the last two weeks together, and in those two weeks I will do my best to condense most of - or let's say the most interesting things - that I've learned in the past two years. In about three hours of lectures you will get to see and learn everything that I struggle to understand over the past two and a half years. This is going to be a bit dense, and short disclaimer, today we will not really talk about networking. Well, actually, not at all. It's going to be a teeny tiny bit of networking part at the end of today's lecture, but today we're mostly going to cover context.

You might wonder why I'm going to spend so much time in giving more general context about sustainability. I think you will understand over the course of today and mostly next week that we need to set the frame of reference to understand what we are talking about when we talk about sustainable networking.

So, let's dive in. What does sustainability mean for you? It's a trendy word. What does it mean? What does it correspond? What does it evoke?

[Proposals from the class]

Something that doesn't harm the planet.

No wasted resources.

The 17 sustainability development goals.

Repairability.

[/Proposals]

So, you're all right. Why are you all right? Because sustainability is a very broad term. I don't want to say vague, because it has a definition, but it's a very broad term. It encompasses many different aspects.

And here's one of the many found definitions, one is the most widely cited nowadays. Sustainability is about meeting the needs of the present without compromising the ability of future generations to meet their own needs.

In this definition, we find a number of the things you mentioned. There's the idea of we should not waste, we should be able to maintain our current standards of living over time, and we should do so in a way that doesn't prevent others in the future to do that.

Note: The [ XXX ] on the slides are references. All links are available in the References section at the bottom of this page.

All right. What does that mean for us? So there is a tiny bit of networking. Something you might not realize for you living in Switzerland is that internet access is far from being universal, still today. The plot stops at 2020, I think now we've passed the 60% mark, but worldwide we're still far from having every human being connected to the internet, even though here in Switzerland we are getting closer to 100%.

Another point that relates to this question is that the electricity demand, the electricity that is needed to power this infrastructure, is growing over time. This is my lifespan. I'm not that old; electricity demand roughly doubled in my lifespan. When my daughter will be teaching here at ETH, the same lecture, 30 years from now, maybe it has doubled again if things keep going the way they are. And this is somewhat a problem.

That's the worldwide electricity demand, it's not just for networking. If you are interested in networking more specifically, it's really hard to get estimate of the global consumption of systems, but we are trying our best. The best figures we currently have are from 2022 for data centers and telecommunication networks (encompassing both wired and wireless communications).

You can see that the numbers are in the same ballpark, and those numbers in TWh correspond to about 1 to 1.5% of the global electricity demand, each. This is just to give you orders of magnitude, try to help you understand the magnitude of the numbers we are talking about.

1 to 1.5%; it doesn't sound like too much, but let's put that into more context. Ireland (which, as you may know, has a very favorable tax system) hosts many data centers. In 2022, the amount of electricity they consume equal all the electricity used by cities. It is a lot. The numbers for 2023 are even higher, but I didn't find a nice graph. In the Netherlands, things were on a very similar track, so the Netherlands actually banned any new data center to be constructed on their ground because the electricity demand is getting too high.

Because those numbers are getting higher and higher, the industry is getting creative. We are starting seeing more and more big companies that are saying, "Okay, you don't want to give us the electricity? If we cannot buy it from you, we will generate it ourselves." The growth of electricity demand seems to be never-ending.

In this lecture, I will talk only about electricity, but it's important for you to understand that sustainability is not just about electricity. There are a number of other dimensions that matter, that matter a lot. Researchers have identified nine of them; those are known as the planetary boundaries. They try to quantify different dimensions in which, if we damage things too far, then we can go back.

Think of a tree or a plant, if you cut a leaf out, it's not that big of a deal. My daughter likes cutting leaves out bushes, and I say "No, stop doing that." Yes, it's fine... it's fine until you keep doing this, and you take more and more leaves out, and eventually the tree dies because you've taken too much! That's kind of the idea here. So we are crossing more and more of those boundaries, it's time.

If we want to avoid compromising the future, so that my daughter can teach this lecture 30 years from now, we probably need to change a thing or two.

So after this introduction, I want to make clear that I've done my best to provide the most accurate, or at least the less inaccurate numbers in those slides. Take everything with a grain of salt. As we'll see today, it's extremely hard to know what we're talking about. It's very easy to find numbers in news outlets or TVs or whatever sort of media that are very catchy, but underneath, there are a lot of underlying assumptions that are most often not clearly explained, but have a huge impact on what those numbers actually mean or imply.

So be critical! [Does not compute] is a short paper (three pages long) that explains things in a lot more detail. I strongly encourage you to have a quick read of this.

So what is the carbon footprint of one hour of streaming Netflix? It's going to be our threat for the first part of this lecture today.

I'm not going to ask you to guess a number, because you have no idea about the orders of magnitude, but let's try to play a game with a comparison. We're going to try to estimate the footprint of one hour of streaming as the number of times you would boil a one-liter kettle.

It's about three boils.

How do we compute this? How do we get this number of three? Well, essentially you can estimate the amount of energy that is representative of one hour of streaming Netflix (we'll deep dive today into what that means) and then use a carbon intensity value that was the average for the UK for that particular year where the report was made, and you get the equivalent of about three boils. Okay. So now we have a number.

How? How does such a number get to be? How would you go about computing such a number? When the report was released, it was quoted in so many articles and then extrapolating by the number of hours and you have the total footprint of Netflix which is huge... okay. But how do we compute this?

So today we're going to cover the following points. First, we're going to look at how we can actually measure this, what does that mean, what are the methods to get to those numbers. Then: What do those numbers mean, in which context can we trust them, or what do they tell us. Finally: Is it really all about technology or there's something else to it?

Let's dive in.

First thing first: I know you're all ETH master students (or most of you). You might fill offended, but experience has shown that this reminder is not completely useless.

Energy versus power. This is not the same thing. Power is the rate at which energy is being consumed. So if we make an analogy with distance and speed, energy is the distance and power is the speed at which we cover the distance. The SI unit for energy is joule. We often use kilowatt hour: it's not kilowatt per hour, it's kilowatt hour. Power multiplied by time.

A joule is one watt second. One watt of power delivered during one second. Consequence: Power is not something that gets consumed, because it's a rate. You can't consume rate, you can consume only energy. We say that power is drawn.

Second point: something that you might not know. We often talk about "carbon" or "carbon footprint". We're not only talking about CO2 in this particular case. In fact, "carbon" is often used as a general term to refer all types of greenhouse gases' emissions. And so this is why you sometimes see gram of CO2 equivalent, or gCO2e, or gCO2eq. All these are the same thing. And it talks about the carbon equivalence.

Carbon equivalence is the way we quantifying the greenhouse effects of different gases. For example, methane, which is another greenhouse gas, has the same warming effect, or is estimated to be the same, as about 84 tons of CO2. So if you would emit one ton of methane (don't do that), it would be counted as 82 tons CO2 equivalent.

Very often, we just say "carbon." Just keep in mind we're not just talking about CO2.

Some more terminology: we hear sometimes clean energy, green energy, renewable energy. Those things are not exactly the same, though there is some overlap. Typically we talk about "clean" for things that do not produce or do not emit carbon. So nuclear would be one example of "clean energy source," at least during the production phase. It produces other waste, but not carbon. "Green" is something that comes from nature, so hydropower or wind. "Renewables" are things that do not expir, i.e., there is no end to their potential (although we could debate that for solar there is a theoretical end...).

Those are not mutually exclusive. There are some overlap, and all this is not really formal. It's just for your understanding that those things are not exactly the same.

First thing that you must understand, if you don't know this already, is that - today - producing electricity emits carbon. It's a fact that you must known and accept. Although there is more and more clean energy, most of today's production comes from oil, gas, or coal.

How much? As of 2022, about 60%. 60% come from carbon intensive, 40% from clean sources using the terminology from before. You need to be aware that the majority of the electricity we consume or we produce today comes from carbon-heavy sources.

Switzerland is an outlier here, clearly. In Switzerland, we have a tiny bit of oil, quite a fair bit of nuclear still, I don't know for how long. But you need to keep in mind that the hydropower is kind of bounded. We've built dams everywhere it makes sense. So we can not scale this up so much more. We can still scale up wind and solar, I suppose. The point is, if we want to keep pushing the production of electricity, we'll have to find some else than hydropower.

Now, we talked about the electricity production. What about the consumption? What consumes electricity in systems?

You need to differentiate two important phases. On the right side, you can see what is known as the life cycle of a product. You need resources, then you need to process them, you need to manufacture the device to send them where it needs to be used, then you have the use phase and the recycling phase. Let's consider only the use phase for now.

In computer networks, what consumes electricity? Do you have examples or ideas in mind? What are the physical things that demand electricity?

[Proposals from the class]

Leakage current in circuits. That's kind of like the waste part. It also consumes electricity to flip the transistors, which is the primary use of electricity in computing units.

The transmission. Yes, over wireless or wired medium. If you transmit over an optical link, you need to turn on the laser.

Cooling. Yes.

[/Proposals]

Those are the primary ones. Speaking more specifically about the type of hardare, we have the routers, which are the main units in computer networks. The transceivers are connecting the different routers (known as "the links"). Something you might not know, which are the optical amplifiers that are along the way, but also all the infrastructure that is required to cool those units.

In this life cycle, we typically differentiate the use phase from all the rest. The use phase is associate to the operational costs, and the rest is encapsulated into what we call the embodied costs or footprint. It is sometimes called "embedded" rather than "embodied;" it's the same thing.

The embodied carbon is very important because it corresponds to everything that you would typically not think about when you you plug in a machine. It's all the rest of things that happen.

So how do you think they relate? If you look at consumer electronics, your laptop, mine, your smartphone, the embodied, so here it's called production, strongly dominates.

I let that sink for a second. So it's good that your phone is very energy efficient, it doesn't consume a lot when you use it. That's great. I'm super happy about this; you can't even see the use phase on the graph becuase it's so energy efficient.

But you should not forget that most of the footprint, almost the entirety of the footprint of your smartphone, come from the manufacturing of it, the extraction of the material, the making of the phone, shipping to your place, and so on.

In networks, things are slightly different. The balance shifts a bit. This is the same type of data as before. The numbers are different because the sources are different, but what is important is of magnitude. For networks and data centers (which means primarily servers), the operational dominates.

Now you may think that "Why is it that my smartphone is so less efficient at being manufactured?"

Actually it's the other way around. It's not that routers and servers are manufactured better; they just suck up so much more energy!

They're always drawing power. A lot of it, all the time. And so, in proportion, it looks like the embodied is so much better, while actually this is the operationak that is so much worse.

Another dimension that you need to understand and keep in mind is the difference between two types of reasoning that you may have when thinking about the footprint of something, which is attributional versus consequential. The purpose of those two reasonings are very different.

In an attributional reasoning, you ask whether something is attributable to you, am I responsible from this footprint?

In a consequential reasoning, you ask "What would happen if ... ?" For example: What would happen if I come to class today versus watching on Zoom? So here you're not so much interesting into the total cost of coming to class, but of the delta in between the two choices you can make.

So let me give you a more concrete example. Let's say we have four people that have to drive to go somewhere to a meeting somewhere in Switzerland and because they are new here in Switzerland, they don't know they should take the train.

They rent a car fitting four person. You do an attributional reasoning and want to know what is the footprint of each individual going to that meeting.

Once you have the footprint for driving the car, the most sensible idea would be to divide it up equally between the four participants. So the footprint for of each of them going to that meeting would be one fourth of the footprint for the car. Sounds reasonable.

Now imagine that a fifth person joins and it's a small car, because whatever, and so now you need a second car to fit the fifth person.

Now if you apply the same reasoning, well, you would allocate two fifths of the cost of one car to each person, right?

Here you need to understand that what people do is exactly the same, but because they need more resources, the allocation that each of them get is going to get bigger.

That's an attributional reasoning. Now let's look at the same story with a consequential reasoning instead.

Now we have these five people, but the need of the second car is because we have a fifth person joining.

We can analyze what would happen if that person would instead stay at home and join the meeting on Zoom. And so here you're not so much interested in the total footprint of those people attending the meeting, but what would happen if one of them joins remotely. And then you can say, okay, I save one car, but it costs one Zoom session. And then you can see what is best. In terms of carbon, it's pretty clear, but there are also other factors you might want to consider...

Good. So the consequential reasoning is really there to try to weigh the pros and cons of different choices you can make. That's its only purpose. It doesn't try to quantify accurately the total footprint. It only cares about comparing A versus B.

Now because those two reasonings are different, they serve different purposes.

The attributional reasoning is primarily made for accounting. Importantly, it has no way of predicting what would happen if something changes.

Whereas the consequential reasoning is designed for this; it's designed to say A versus B, what happens if I do things differently?

So the consequential reasoning is what you need to predict, whereas the attributional reasoning is what you need to distribute responsibility among stakeholders.

Now, if attributional is for accounting, you need a way of saying how this accounting should be done. You need some rules for this accounting to work. And this is what the GHG protocol is about.

Show of hands, who knows what or heard about the GHG protocol before? Okay, no one, really? Okay, I'm mildly surprised... So you all have something to learn, great!

The GHG protocol is a way of classifying responsibility of carbon footprint. It works by dividing emissions into scopes: scope 1, 2, and 3.

Scope 1 are the emissions for the fuel you burn yourself, being coal, being gas, being whatever.

Scope 2 is for the electricity you use that was produced by somebody else. I've told you before in the intro, production of electricity has a carbon footprint. That comes into scope two.

And scope 3 is all the rest. What does that mean? It means all the emissions that come from the people you buy your product from, the people you're selling your product to, and so on.

Here's a way of visualizing the scopes. Let's say you are ETH and you want to do your carbon footprint analysis: you are here, in the middle. You maybe have facilities that run on gas for heating; that will count into scope 1. You are also buying electricity to power my laptop right now, or the beamer; that will go into scope 2. The upstream part of scope 3 would be the workers coming to work, going to conferences... Since ETH is not selling anything, we don't really have anything downstream. But you get the idea.

What you need to understand is that, for the same product, its footprint would fall into different scopes depending on how you get it.

Let's say you want to make a cup of tea, or coffee, or whatever you like. If you're in a hut, in a cabin somewhere in the mountains, you will probably have a wooden stove. That stove will burn fuel to heat the water to make your drink, so your coffee's footprint would fall into scope 1 in this case.

If you are at home, you may be using a kettle that run on electricity. That electricity has a footprint but it will not count into scope 1: it would count into scope 2.

If you go at a Kiosk and you buy your coffee from the machine, you do not pay for the electricity; somebody else's does. That somebody also paid for the cup (if you took a disposable cup from the Kiosk. Next time, bring your own!). Those footprints are ultimately related to the same product - a cup of coffee - but now count into scope 3.

Scope 3 is really tricky because it's really hard to set the boundary of what you count in and what you don't. Sometimes it might seem intuitive to identify who are my customers or not, and so on. But sometimes, it's really hard to say where do you stop.

Think about the footprint of a video call, similar to how we started with the footprint of an hour of streaming. What do you count into that? (I'll give you the answer a bit further down.)What do you count in? Would you count your laptop? You need your laptop for getting into this video call, but you don't need your laptop only for this. But still, if you didn't have the laptop, you couldn't do the video call. So in or not in? Partially in? How much then? It's not trivial.

If you followed what I said before with the scope 3 being everyone that you buy from and you sell to, and if you take that reasoning further, you may realize that your scope 3 is somebody else's scope 1 and 2. Because you bought your thing from somebody, and that somebody spent the electricity.

In my coffee example, you buy your coffee from the Kiosk; the Kiosk spent the electricity for your coffee. So it's scope 2 for Kiosk, but scope 3 for you. But if we all count, then we double count.

Is double-counting good? Is it bad? It depends what you're trying to achieve.

The GHG protocol is setting a framework. It says, "Okay, here's one proposal of how we could classify the different costs." That's about it. There's a bit more to it, but it's essentially the only thing the GHG protocol does.

Then what you need is some sort of implementation of this framework for a given industry. One of the most interesting guidelines I know of is called DIMPACT. It was precisely designed to "Okay, we take the GHG protocol and we clarify what should go in or not for online entertainment and video" (which, you might have noticed, relates to the example I picked for this lecture).

There is a collection of such implementation guidelines that clarify what we need to count for transport, for entertainment, for food production, and so on.

There is another dimension that you need to keep in mind - I know, it's a lot of them.

This is the difference between top-down and bottom-up approaches.

So let's say you would like to know how much electricity does emitting a ChatGPT query consumes. Sensible question to ask. How would you do this?

Well, if you do this in a top-down way, this is how it could look like. Let's pretend that you can know that your query you're emitting right now comes from that particular data center here in Zurich, and that data center has an energy consumption of 100 units of energy. This is what you can measure. The data center is plugged into the grid. This is the entirety of what the data center consumes in terms of electricity to serve this ChatGPT service.

Now of course, in that data center, we have many servers. Let's pretend there are three of them. If you don't know any better, like in the car example we had before, the only thing you can do is dividing up the cost into the different servers. Each server will run different processes to serve multiple GPT queries in parallel, and again, if you don't know any better, there is no other sensible choice than to equally divide up.

It means that as we go from the top - where we measure, this is what we know to be a fact, that's the electricity we consume - to the information that we are interested in at the bottom, we lose information. In the middle, we have no idea what's going on. At the bottom, we end up with a smooth signal without any texture: all queries are the same, all processes are the same, and if you run ChatGPT but I run a good old Google search, then at the end of the day we get accounted the same, because we don't know how to differentiate.

It sounds silly, right? But that's what a top-down analysis does. It's an extremely common way of analyzing data.

Top-down analyses are very good for auditing, because for auditing, we don't care about breaking things down. What we care is what is the total. So every time you have an audit, it's made with a top-down analysis, almost always. But it doesn't really tell us anything about which process to optimize in order to reduce my energy demand.

On the other hand - you see where this is going - if we follow a bottom-up approach, then it's different.

Let's pretend that we can know exactly the amount of energy that is used by each process. The truth is, it's not so easy, but let's pretend we can. Then you could say, "Okay, I have all my processes running right now, and I aggregate as I go up. I aggregate per server, and then I aggregate per data center." Why do I find 20 instead of the 100 I had before? It's the same data center. Nothing changed!

There is information loss, but the other way around. Why? Because we never count everything.

When you count bottom-up, you know what you count, but you don't know what you forgot to count... We had the example before: when I asked what consumes energy, the first thing that came to mind - you're electrical engineers - was leakage. Yes, of course, there's leakage. Okay, let's count the leakage. We count the transistor, yes, okay, we do this. And then? Oh, yeah, there's also the cooling! Okay, so we need bubbles for cooling as well...

There is more. There's always more. There is always more things that consume any energy that you forget to count for. It's almost inevitable. So when you do bottom-up, you have the higher granularity, you know what processes are consuming more, so you know that you probably want to optimize, try to optimize this one, and that other one probably doesn't matter that much because it consumes much less.

That's not going to be good for our counting. It's not going to be good for our counting because you're going to say, "I sum up the numbers and I get 20" and then you look at your energy bill, and you're like "Wait! Why does my bill says 100??"

Both top-down and bottom-up are useful, but for different purposes.

So I've talked about many different dimensions, I know it's a lot. So here's the recap on one slide. When you talk about sustainability, you need to think about:

Are we talking about operational versus embodied? Do we care about the operational phase, or the rest, or both?

Am I trying to do some attribution or am I trying to do some consequential analysis? Am I trying to estimate, would it be good if we did X, or am I just trying to get a sense of what is my total responsibility as an individual or as a company?

Am I looking at something that's more top-down or bottom-up in terms of analysis?

What scopes are being included?

I think by now, you start getting a sense of why we don't have time to talk about networking today. Because if you don't understand all this, then for any of the things we talk about when we talk about sustainability (not just in networking, but in any aspect of your life), it's very easy to be misled.

I see a number, numbers are great. I showed 55 gCO2e before.Great, now we know! But what is it we know exactly? ... Not so clear. So what is it we know?

Oh, sorry, I forgot this parenthesis. So you have all those different dimensions, and you need to clarify in your mind what is it you're talking about.

If you want to improve your operational footprint - here we're talking about only operational, so forget about the embodied - we can look at the carbon efficiency of your system.

That efficiency is made of the energy you use divided by the amount of tasks you do, multiplied by the efficiency of the power conversion (that is, how much energy do you need to supply compared to the energy you actually used for the task itself; there's always losses, we talked about leakage before, this is one example of this), and multiply again by how much carbon you spent to produce that energy.

So you have three ratios. If we forget the last one, we get energy efficiency. Iif we include the last one, we get carbon efficiency.

The first term relates to utilization and what is known as power proportionality, or energy proportionality. We're going to talk a lot more about this next week, so I'm going to go quickly today.

The idea is that when you have a device turned on that you don't use, you still pay some energy costs. Like my phone is off over there, it still consumes energy even though I'm not using it, even though my utilization, so to say, is zero, the power draw is not zero.

The second term is what we, in networking, usually call power usage effectiveness, or PUE. Somebody dubbed this one the "pointless use of energy", which I found to be kind of an interesting play of words. It's kind of that: How much energy do we loose compared to what it actually needed to perform our task?

If you think about a data center that is connected to the grid, and if you draw 15 kilowatts at a given point in time, you may have 5 watt for cooling, so, a third (which is a reasonable outdoor magnitude by the way) and 10k for actually running the servers: you loose 33% here. It leads to a PUE of 1.5.

Then, if you continue down the chain, in your server you may have 5k of losses that come from the lack of proportionality that I showed just before, and 5k to actually running the app. If you sum everything together, your app itself only takes 5kW to run, but you are consuming 15kW in total.

PUE was introduced reasonably recently, and since then things have improved significantly.

When it was introduced, values were around 2.5, which was awful! Think about it: it means you're spending more than twice as much in overhead than doing actual stuff.

By now the hyperscalers - so here's an example for Google but the others did the same - managed to get the PUE fairly close to 1. That means we have less than 10% of overhead. Getting further down is going to get harder and harder, as you might expect.

And finally, the third factor is what is known as the carbon intensity. So website is called Electricity Map.

[Live demo]

It's a pretty great service that shows you live what's going on in the grid, and gives you the average carbon intensity for a given location. In Europe, locations are mostly country-based.

So you can see that things are different depending on the places. If you click on a country, you can get a breakdown of what is the current energy mix.

In France, as you would expect if you know anything about the French power system, most of the energy comes from nuclear, and so it's rather green. Whereas in Switzerland, it's a bit more of a mixed bag. You have solar, you have hydro storage, and so on.

If you go to Poland, the carbon intensity is more than 10 times as high as Switzerland. Right now, it would be 15 times as high as what we had before, roughly. This is because most of their energy comes from coal sources.

So that's the live right now. You can zoom out, and you have other locations on the planet. You can look at the different states in the US if you care. And there's some other interesting things.

This is the live view, but you can also go back in time. So let's do this while viewing Europe. As I go back in time, the colors change. There is no miracle. Poland is always driven by coal, so Poland will be always bad. Whereas Switzerland will be mostly good, because usually they have a clean energy sources available most of the time.

Very interesting website, which of course also has an API that you can query to get information about programmatically and so on.

[/Live demo]

Here is a view of carbon intensity over time for three different cities in the UK. So here you have the hours of the day. So you have 24 columns and 52 rows. So this is one year. And what you can see is that you have places where, like in London, when the wind blows, things are OK. But when it doesn't, well, it's not so OK.

And it varies per week of the year more than time of the day. You have places where it's kind of always good, and you have places where it's kind of always bad.

The consequence of this? If you think in terms of consequential reasoning, you can quantify what happens if I run my compute jobs in Dublin or in Scotland. If you were in the consequential reasoning mode, you could use the delta in carbon intensity to quantify the impact of that choice. Assuming everything else stays the same (note: it does not), I take my data center, I put it in Scotland: Ta-da! This is how much I would win. Again, it would not be about "What is the total footprint of that data center?" but "How much I expect I would save if I would move it to Scotland?"

This breakdown is useful because it identifies different sectors of the whole system where you can hope to make improvements. If you improve any of those terms, then you will improve the total.

We're going to talk a lot more about "device efficiency" next week and I will mention briefly the other two sectors at the end of today's lecture.

In the first hour, I provided a lot of context, and I've presented this yet mysterious number of 55 gCO2eq per hour of streaming Netflix. So now let's try to unpack this, and understand where it comes from.

It comes from somewhere, and I'm not here - I'm going to say that right now - I'm not here to bash Netflix. The reason they're here on the slide is because they are one of the few that actually give out the numbers and explain how it comes from. So we should be thankful to companies that actually make an effort to be transparent on that regard. Let's try to understand where these numbers come from.

If you watch Netflix, you're using somewhat this infrastructure. Here again, use phase only; so operational footprint only. We have data centers and something called CDNs, where the movies or series are being stored. Then you have the actual networking. And finally you have the end-user devices; whatever you have at your place, or when you're moving around, the device you use to watch your stream.

That makes up three categories. Let's consider the energy use of those three categories: What do you think is the share of the network here in the middle, in percentage? Three guesses, at least. You cannot know. So no rush, no pressure, it's random guesing time.

[Proposals from the class]

80% 60% 1%

[/Proposals]

It's around 10%. Okay. Now what about the storage, the data centers?

[Proposals from the class]

80% 40% 20%

[/Proposals]

We had the right answer before. That's 1%. The energy cost for storing movies is nothing. I mean, not "nothing." It's rougly 1%. Transportation is about 10%, mostly cellular, but wired also costs.

Most of the energy by far is consumed by your phone, your TV, your laptop, at the time where you actually watch. Okay? I did not expect that when I prepared this lecture.

But it is something very important to understand: When we talk about streaming Netflix, in the operational phase, when you actually stream, almost all the energy goes to the device you hold in your hands, you have on your laps, or you're hanging on your wall.

Now if you look at the GHG protocol, into Scope 3, there are different subsections. One is "use of sold products." So if you think about it, your laptop is a device that was sold to you, but not by Netflix. Netflix is not responsible for the device you choose to watch their stream on. So they say, "We account that, in our scope - what we count as our responsibility - we're gonna have the share of the data center on which we store the content, the share of the network we use to deliver our streams, but the rest is not on us. It is outside our scope."

Remember that the accounting rules are somewhat arbitrary regarding what you choose to count in or not is, you know. A subjective choice. At least, Netflix transparently say "This is our methodology, we report our numbers, this is how we count in, and we leave that other stuff out." They leave out the 10% of the network and the 89% of the end-user devices.

Now interestingly, the 55 grams I reported before does actually include end-user devices. So now you're saying like, wait, is it "in" or "not in"? Well, in fact they do two accountings. One with, and one without; which arrives to footprints in megatons of CO2 equivalent. They take the one without and say "Here is the share we are responsible for and that we will work hard to make smaller over time." Then they also present the total number with end-user devices counted in and say "this is the total footprint, and even though a large part of it is not ours to improve upon, the electricity is still consumed, and that total number matters." That total, when you divide it up by number of streaming hours, is what lead to 55 gCO2eq/hour.

Okay, now I'm not going to play the game again, but if you consider Netflix and look at all their activities, the production of the movies and series actually strongly dominates the streaming part.

And if you think about it, every time they produce a show, they need to fly in people to the location and then drive to the thing, set up the camp and shoot the movies and so on. This largely dominates everything else. In particular, it largely dominates the streaming part, which is what we think about when we think about streaming Netflix. At least, this is what I was thinking about; I was like, "oh, the Netflix report will be a great case study for the AdvNet course!"

In fact, well, the network part is, you know, just this. So if you think about it before, we said storate is 1%, transportation is 10%. So the right box is about 11% of the total energy footprint, and that is only 3% of the total carbon footprint of Netflix.

Finally, it's very important for you to understand that those numbers are attributional.

How those numbers get computed? Netflix looked up everything they spent, all the flight they needed to book for their teams, all the diesel generator they had to power to heat up when production and so on, all their buildings, and then you divide it up by total hour of streaming served, and that's 55 grams.

This is how you come up with this number. But that means that if you decide NOT to stream tonight for one hour as you usually do and do something else instead, you will NOT save 55 grams of CO2. Why? Because it would be correct if this number was derived from a consequential analysis, but it isn't. It's attributional. It's meant for accounting. It doesn't tell you anything, literally nothing at all, on how much you would save if you don't watch (or if you watch more).

It is extremely important to understand. This 55g number (and others that were much more wrong) were used to "shame" people into not doing X, or propose actions that make no sense. Streaming more or less may have an impact on the Internet footprint (we will talk about that next week) but it has nothing to do with those 55g. It's very different, and it's important to understand the difference.

That's the theme of today's. I'm really hammering this in, but that's THE one goal for today's lecture: you must understand that numbers need to be put into context. We need to understand how numbers are derived and for what purpose, because this purpose dictates what you can say or not say about the number, what the number enables you to understand about the system.

This 55g number is useful. It is very interesting to have, but it's useful for Netflix to look at how their total footprint evolves over time. They did this in 2022, they did this in 2023, they will do this again. More and more companies are now compelled to do this sort of analysis. It's being compulsory in the EU for data center operators now. So then, you can look at your total footprint, see how much useful work you actually do, and then look at the global efficiency of your system overall. That has very little to do, usually, with the marginal footprint of the final product, and that's particularly true for streaming.

Let's move to part two. Can we trust the numbers? What do you think? I don't know if you ever came across this: There are these sorts of statistics about when you have a question in a title, the frequency at which the answer is no versus yes. Have you ever seen this? I don't remember the stat. It's almost always no.

Here it's not really no, but it's not really yes either. If you had followed what I said the first hour: it's yes if you know how the number was derived.

Give you a second to read this. This is what we're trying to avoid, okay? A lot of companies report numbers that are not exactly wrong, but they are not what they imply either.

How many of you heard the term carbon neutral before, okay? How many of you heard about net zero before? How many of you know whether there is a difference between the two? There is. It sounds kind of like neutral zero, net zero, like it sounds kind of the same thing, and it isn't, but not at all. So what does it mean?

Those are convention, by the way, to provide a common understanding when we talk about things.

When Google say "we're carbon neutral," or Amazon say "we're carbon neutral," or anybody else, what they mean is this: It means they are offsetting the emissions they are responsible for by some other means. Usually, this is buying what is known as "carbon credits." So, that means you emit whatever you want, as long as you buy credits for the same amount. That's what carbon neutral means.

There are a couple of issues with this approach. In practice, when people do this, they often count Scope 1 and 2 only. So only the fuel you burn, or the electricity you buy. Say I buy so and so much. In my country that results in so many grams of CO2, I buy credits for the equivalent. In the digital sector in particular, scope 3 often heavily dominates; so digital sector is more in this 95% range than 70%. It's a complicated topic, I don't want to get into this, it would be a whole lecture in and of itself. Bottom line: it is problematic to only count scope 1 and 2.

And typically those offsets, so those carbon credits, tend to be of low quality. I'll explain a little bit later what this means.

Net zero on the other hand, by definition, and this is a more formal definition, includes all the three scopes and has clear reduction targets. Net zero is not just a state, it's more of a trajectory. It's something we haven't reached yet and we aim to go to.

In the definition of net zero, it is clearly emphasized that the priority must not be the offsetting (like with carbon neutrality), but must be the reduction of the footprint, so that you have little left to offset. You always have some footprint. You always need to spend some and to buy some electricity at least. But the point is to first reduce and then offset what you have left. And if you do, those offsets need to be of high quality and high quality in the sense means that it needs to relate to removal of carbon that has been previously emitted through reforestation or through different sort of means. It's a much more narrow scope of what you can offset.

Another important thing to keep in mind is that it does not help to selectively buy the green energy. Today, Amazon is the largest buyer of green certificates worldwide, by a large margin. But if we pretend that this pie here is the electricity demand, or we have green here and carbon heavy there (Eyeballing the 60% of carbon-intensite share of electricity production from the intro), what it means is that AWS, Google, and Microsoft say "we take those shares and you get the rest." So cool, they can claim that they run on full green electricity. That doesn't help the total. But this is what's happening.

This is what's happening when companies say we run carbon free service. On every single Amazon website you can find and marketing docs saying "Come run your workloads at AWS, it's 100% carbon free!" What they mean is this: They buy a boatload of certificates for this. Is that really what you want? Probably not.

Time for more details about what those certificates are. The terms we use in Europe - it's a bit of a mess because we use different terms in different parts of the world - is energy attribute certificates, or EACs. The motivation behind them was the following: We want to foster the development of renewable. We want to encourage people to deploy more solar and more wind. Sounds sensible.

So the idea was we're going to ask people to pay extra for their electricity, and that extra money - a sort of tax - will be used to foster the deployment of new solar farms and stuff. So the idea that with this extra money, with this premium, then the producer will be paid more than what the electricity costs so that it will encourage them to deploy more renewables.

It sounds sensible... until people are starting to need to make some accounting and wanted to say they are carbon neutral.

In accounting there are actually, yet again, different subtleties. There are two important approaches that are what are called "market-based" and "location-based".

The market-based definition says that you can buy credits wherever you want. It's an open market, like for any other resource. So you have huge solar farms in Australia, for example, and they say we produce so and so many watt hours of electricity per year, and you can buy credits for those. You can buy them wherever you are. So you're in Germany or in Poland, as we saw before, you can buy those credits. But you're not consuming the electricity that is produced in Australia because this electricity is not being carried back to you... That's not how it works.

On the other hand, you can have location-based accounting where you say you can buy only things that are produced close to you, to which you are actually being connected. So if you want to buy, let's say, Swiss credits, they need to come from renewable power sources within Switzerland. And if you're drawing power from the Swiss grid, it makes a lot more sense. So if we want to be more honest about the offsetting that we do for our business, then we should do location-based and not market-based.

But now comes the funny part. The folks who produce this electricity, this green electricity in Australia, they sell certificates on the market. That means, if you want things to be fair in the global scale, the electricity corresponding to those certificates should not be used by anybody else. Because you've "sold it" (more precisely, you sold its greeness) to somebody else offshore.

Needless to say, this electricity is not being wasted, which would be stupid. And so consumers in Australia who buy that electricity say "we do things right, we do location-based offsetting!" You end up having certificates that are used in multiple places at the same time.Check out the [ Green certificate ] article for more details about this.

There is another issue: Those certificates tend to not have any "additionality." What I mean here is this. We aim to encourage producers to install new renewable capacity. But it's not really what happens in practice, at least not always. Why? Because in Switzerland, for example, we had for a very long time a lot of hydro power. It was there. It was there for a long time because it makes sense. And so when the system came into place, well, we converted this amount of electricity into... "Oh, now we have these carbon credits we can sell to people. Great. Free money!"

But that money did not serve to build new dams because we cannot. So in many cases, we sold credits for power that would have been generated anyway. So if you take back the energy pie chart image: regardless of how big the green part is, Amazon now claim they consume green (though, in fact, they consume anything)l and we are not incentivizing the deployment of more green. So it makes no sense, zero progress, only greenwashing.

Another subtlety is that when you talk about carbon neutrality, it depends if you try to be neutral on average over the year or at every point in time. Because, of course, if you think about wind and solar, you cannot generate all the time. Solar is kind of bounded half of the time, maximum. And so if you generate a lot of solar in the summer and nothing in the winter, you buy excess credit for the summer, you average it out over the year, ta-da, you're carbon neutral. Even though, in the winter, you run a lot of coal to get your electricity. So that obviously doesn't make a lot of sense. And this makes a lot more. You're trying to offset for every hour your consumption at every hour in time.

Naturally, people have realized this, of course. And so there has been other proposals to recapture the original purpose that we had for those green certificates.

The main lever that is used today is known as "power purchase agreements", which is something that existed for a long period of time but has been renewed in that context. The idea is to say: "I'm a company and I commit to buy so-and-so amount of energy to you for the next five or ten years."

The benefit of this is that you're not buying just when you consume; you commit to buy at a certain price, which is usually fixed. The idea here is that this gives electricity producer a lot more visibility. They're going to have this regular cash inflow for the next ten years, so they can invest that money into building new capacity.

Now what happens is that we're starting to have those power purchase agreements, those PPAs, with conditions for additionality. Because in themselves they don't guarantee this, but now we start having some accounting that says okay, if you want to count that PPA, it needs to come with some commitment from the electricity producer that this electricity comes from a newly built power plant, or that newly built power plant is being constructed from the money that comes from that agreement. So we're trying to get back to the original purpose of fostering the deployment of renewables. And it is working to a certain extent.

Again, it's important to clarify what we're talking about. Are we talking about net zero, are we talking about carbon neutral? It's important to have metrics to measure those things.

This is an adaptation of a very famous quote that I'm sure you've heard many, many times over. You cannot improve what you cannot measure.

But also, if you measure the way you want, you can reach whatever conclusion you want. We'll talk about this a bit more next week.

The proper definition of how you measure matters, even if all the rest of the scope is correct.

This image kind of captures this. In that image, you have a bunch of people who look at the same object - an elephant - but they look at it from different angles that would be different metrics. In their own perspective, this object is still always the same, but if you look at it with one very specific or narrow metric, you can reach whatever conclusion you want.

So if we want to really know what we're talking about and we want to progress, we need to have standards to help us define those metrics and those standards need to be followed.

Here is one example. It's a fairly recent, I think it was from last year, metric that was proposed and is now an ISO standard (one of the most reputable standardization bodies).

This metric, the SCI, or "software carbon intensity", is a formula that works as follows. In the operational phase, you consume so much energy when you run your program (E), you run in a place where you have a certain carbon intensity (I). E.g., in Switzerland, we saw before that right now it is about 56 gCO2eq/kWh. Then you need to add some value for the embedded cost (M), and that's a fixed term. Finally, you scale this per unit of work that you do. So, if your software is an LLM, a unit of work is a query. If it's an email client, then per email sent.

The SCI is one way of standardizing how we count the footprint of something.

I leave to you as exercise to think in which cases this metric is useful. For what kind of reasoning can you use it? What is this going to tell you? What can you do with it?

So now we have metric, we know where we want to go, we want to go to net zero (not carbon neutral).

So that's net zero (right part of the graph), that's where we are right now (left part), and what we want is a pathway that's going to take from where we are right now to net zero. Of course, what we want ideally is something that looks like this over time (left graph), and not something that looks like that (right graph), which looks quite silly, right? In principle, mathematically speaking, this would also be fine. You get to the point you want at the time you want with a continuous function, but that doesn't sound super encouraging.

So ideally we want the left graph. This ressembles the shape we've had before on the PUE graph that is getting lower and lower over time.

So we've got the metric, the SCI or others, we have the target...

... but we are not anywhere close reaching them. That's the CEO of Microsoft.

In 2024, Microsoft published their system ability report, and it looks fancy. It's full of green, it looks great, it's full of green, wonderful!

If you look at the caption, those are the scoped emission of Microsoft. As mentioned before, scope 3 largely dominates. Then we have scope 1 and 2 at the top. At the bottom, we have the carbon retrieval programms.

So you take top part, you subtract the bottom part, and then you get your net footprint. The caption of the table is, you know, what we can see from here to here, overall emission increased by 30% in 2023.

Put into a scatter plot, these green dots are what we said we wanted, and the black is what the data show. So it seems like we are in a situation much more the red curve we had before, rather than what the pathway we said we want. Maybe, magically, at the end of the day we're gonna fall down, but you know, it's not very encouraging.

Again, I don't want to single out Microsoft here. We can say this about Microsoft because they are actually being transparent, publishing those objectively pretty bad numbers. If you keep reading the report they say like, "Oh yeah, but you know, this is a one-time investment for the new AI stuff. It will stop." We shall see... Again, it's not to single out Microsoft; the other players are not doing much better.

Be careful about ratios. In the intro, I talked a lot about efficiency. Efficiency is a ratio, right? It's the total cost versus the total amount of useful work you do. And it is useful, it's useful as a guide, right?

Those are the two graphs I showed before; it's the same data. The left shows the relative energy production, and that looks good! It looks like coal is kind of stable - we would like to see it going down, but it's somewhat stable. But at least, wind and solar are picking up! They are getting bigger and bigger, that sounds nice. But if you look at the absolute numbers, coal is actually going up. It's actually going up, and the total also goes up. Although the share of coal-based production remains constant, the total goes up.

The point is: relative numbers do not tell you the full story.

So when you see a report and you only get a relative number, or only get an attributional number, in the back of your head, you must ask yourself: "Okay, but what's the total?"

Here's another example in a different context: Energy use and energy efficiency of networking equipment.

On the left is a graph I redrawn from a presentation from Broadcom. Broadcom is one of the vendors of the ASICs that are inside the routers. This tends to be the most power hungry component in a router or a switch. Broadcom reports that the efficiency of those chips (in watts per 100 Gbps) is going down, like this nice exponential curve that we saw before. Great! We are doing so much better and better over time. This means, for 100 Gbps of traffic we forward with such a chip, we are spending less and less energy. It is good.

But what's hidden here is that this is enabled by the fact that we are having faster and faster Ethernet speed over time. That comes with bigger and bigger power demand over the same time period. Indeed, 22 times increase in power is smaller than 80 times in bandwidth, so the efficiency goes down, that makes sense. But the total still goes up. It's important because, you know, we have on the one hand improving efficiency, on the other hand reducing the footprint. But in the end, only reducing the footprint matters.

It does not matter how efficient the things are: what matters is that the total amount of resources consumed goes down.

Reducing the footprint is what we need to do. Improving efficiency is just a way to get there, it is only a mean to achieving footprint reduction. By improving the efficiency - if everything else stays constant - we would reduce the footprint.

Sadly, it's not even guaranteed that this is going to happen. In fact, there is a very important phenomenon - the last thing we're going to talk about today - where improving efficiency may in fact increase the footprint.

Let's take a step back from the technology discussion.

There is another phenomenon that is extremely important for you to be aware of: It is known as the Jevons paradox.

Side note: Jevons was an economist who made the observation that when we started building more efficient coal power plants - i.e., getting a lot more power out from burning coal that we used to - it sounded great, but the net effect was that it greatly increased the total amount of coal we burned. This easier access to electricity is what fueled the industry revolution.

Here a depiction of England, where it looked like coal power plants all over the place. You know why? Because all of a sudden it became a lot more interesting to burn coal for energy, and so we started doing a lot more of it. The efficiency was better, it made it more interesting to burn coal for power, which created an incentive to consume more of coal, and in the end the total consumption increased. It is a complex mechanism, but very common and something to keep in mind and be aware of.

The "Jevons paradox" is also known as "rebound effects."

The Jevons paradox is observable in many contexts, and many times over in history.

In the ICT sector, we can see this as well. The years are kind of arbitrary because they depend on where I found the data, but over time we've made a lot of progress. The phone I have today is not particularly fancy, but it is, I don't know, maybe 1,000 times more efficient than the first phone I had when I was 15. And yet, we still use a lot more electricity for our phones now than 15 years ago because everybody has one or two, whereas before, not everybody had one.

Same thing for the carbon footprint: For example, we have a lot more phones. As I mentioned, their footprint is dominated by the embedded cost. So we produce a lot more of them, we change them every couple of years, and every time we do so, bang, you pay the embedded footprint again. And so the issue is not only the sheer numbers of phones we have - after all, they do not consume so much energy - but it's also the cost of manufacturing them that increase significantly.

Bottom line: there is one thing that you must absolutely understand. If it's the only thing, please remember this.

Every time we ask for more services, it has a cost.

If we ask for more stuff, it comes from more resources. The efficiency gain may compensate that increase, but it's easy to understand that if we would have the same efficiency gain while the demand remained constant, then the footprint would decrease. Efficiency gains are always good to have, and they will be good even if we don't ask for more resources. But if you ask for more resources, you will consume more.

This is captured by a term getting some traction today: "digital sobriety."

You need to understand that every single thing you do in your life - not just in your digital life - has a cost. Every email you send, every video you watch, every XYZ. We'll talk about quantification next week. It has a cost. That cost may be small, but it is there. It exists. And you should not neglect it.

And I'm not here to tell you, "Don't watch Netflix," or "Don't send emails," or "Don't store photos" or whatever. Of course you can do that, and you probably should, but be mindful of not doing it without any purpose. Try to do it when it really brings you value.

And of course, I'm not trying to shame you here. It's not just about you. It's also about the entire society. For example, we have a huge problem with data storage. I'm not going to get into the details because we don't have time for this.

In short, we store gazillion of files that nobody cares about. But it's stored, and it's kept forever. Every single file we store has a cost. Remember the Jevons paradox? We're getting so good at spending little energy for digital storage that it does not matter so much anymore. But remember, energy is only the operational phase! There's all the embedded phase.

When you store one terabyte extra of data, that means it's somewhere on a hard disk that had to be manufactured. The cost of this is huge significant. Some people call this dark data, which they found it funny - like dark matter, it is there and pulling us.

Let's summarize.

How do we measure footprint, be it energy or carbon footprint? I tried to explain to you today that there is a lot of subtlety here. There are many different ways of counting, different ways of looking at the question, and importantly, none of them is "wrong." They all have a sense. They all have a purpose. But it's important to understand which purpose they serve so that you can interpret the outcome correctly.

Then I tried to explain that you need to be careful with what you read or what you can see in the media, being social or traditional media. The companies that put those numbers out, they know why they count the way they do. They purposefully choose the angle that tells the story they want. If you don't understand this, you will be manipulated. But if you understand those different analysis angles, you can dissect the numbers and interpret them properly so that you don't get fooled into believing what people want you to believe.

Use your critical thinking.

You're soon going to be ETH graduate students. If you, among any people, cannot understand and decrypt what you read, then I have little hope for society.

Finally, I think it's important to understand that technology is not going to save everything in the end. Even if my research is tremendously successful and I reduce the footprint of networks by 10x, if, in parallel, the users ask for 100x the bandwidth, then there's nothing we can do.

Be mindful of what you consume. Do it, there is no shame in consuming resources and digital services, but try to not waste things. It sounds silly, but - sadly - it seems that it's useful to be said.

Next week, we'll dive a little bit more in the first part here. So how can we use less energy for the same task? In this product of efficiencies, this is the main part where the network comes into play. It has little to do about the power conversion and about the carbon intensity of the electricity.

We're going to talk about this a little bit - very briefly - in ICT context.

Don't quote those numbers. They are from an old source, but I didn't have time to find newer ones.

The point I want to make here is that cooling and lightning and so on represent sizable costs.

So if you want to improve your efficiency, you can do two simple things. You can try to cool less. There's this amazing paper - starts being quite old now - which made the observation that if we increase a little bit the temperature of a server, the failure rate actually does not go through the roof, as the theory says.

We used to cool data centers at 18 degrees; that used to be the standard mode of operation. They found that up to 27-29 degrees, it's all fine. You see no significant impact in the wear and tear of the hardware. But, of course, if you only have to cool to 28 degrees, you spend a lot less energy than if you need to cool to 18 degrees.

This academic paper effectively changed the way we operate data centers today, so that's pretty cool. It's as big as it can get as terms of real-world impact, I think.

Another idea that is being explored more and more is to try to reuse the heat that comes from computing or networking. So in terms of systems that are being explored nowadays, to pump the heat back into growing food facilities, or heating swimming pools, or apartments, or things like this. So we try to reuse the heat, which is a byproduct, and so overall, the hardware still consume the same amount of energy, but that energy is being extracted and reused so that we don't have to spend it the same amount to heat up water elsewhere.

Finally, carbon intensity. What do I mean by application elasticity? Well, I've showed you before on ElectricityMap that things vary over time and space. There are places that are always bad, but there are places where things vary. In places that have a lot of solar, like in California, when the sun shines, which is often but not always, you have a lot of solar electricity. But you have time when you don't, and so it makes sense to try to tune your consumption up and down depending on the sun and the wind.

So this is something that's called as carbon-aware X, carbon-aware computing, carbon-aware networking, carbon-aware whatever you want, and the idea is to either wait until you have green energy, or move to a place where you have it now. So if you think about Google that operates YouTube, there's a lot of work that needs to be done to encode the videos. This encoding can be done anywhere. It doesn't really matter where you do this. So Google is doing a lot of work to try optimizing where and when they do the video encodings for YouTube so that they do this at times where the carbon intensity is low, which reduces their overall carbon footprint.

Similar things are being experimented by Microsoft. For a very long time they had this feature that your workstation would download Windows updates at night, but now they're starting to realize that night is not necessarily the time where you have the lowest carbon intensity. So they are now experimenting with a feature where your laptop will automatically decide to download the next Windows updates when the carbon intensity is low at your location.

So those are kind of like things people are thinking about.

If you transfer those concepts to networking, the idea would be to route your traffic through the greener paths. I put here two recent papers that explore this. We're going to talk a tiny bit more about this next week.