What Exactly Is Green Coding, and Does GenAI Make It Easier? An Interview With Max Westing From the “Green Coding” Project

Did you know there’s an eco-label for software? The ‘Blue Angel for Software’ recognises solutions whose programme code is particularly resource- and energy-efficient. So if there are ways to make programmes and websites more sustainable, why isn’t this the norm? And how does green coding actually work?

Our interview with the ECO:DIGIT project showed just how complex it is to measure the environmental impact of software. This is mainly because software and websites are no longer run solely on end devices, but also in data centres. For users, the advantage is that we no longer have to wait for a DVD in the post for a film night, or connect hard drives containing our video collection.

We asked Max Westing from Trier University of Applied Sciences whether green coding is similarly complex and how programmers can integrate sustainable practices into their work.

Max, how do you actually write software?

It’s quite complicated and difficult to answer that because there are always different levels of abstraction. At the end of the day, you usually program in a development environment where you produce your source code.

Nowadays, this is quite often done using generative AI. However, you are always generating code that is readable by humans but is translated into code that machines can execute. And there are different levels of abstraction—that is, languages that are closer to or further from machine code than others.

A good example of this is Python: the standard implementation is of the computer language C. So when we run Python code, it is translated into C code. And this C code is then, via a few intermediate steps, translated into machine code—that is, into the sequence of zeros and ones that computers can process.

So what exactly are these ‘code libraries’ that you keep reading about?

When you load a code library into a development environment, it is ultimately source code. In this case, however, it consists of objects that are made available to the system and call specific functions. This saves you the trouble of having to re-implement certain functions time and time again.

A so-called ‘library’ is like a standard component that you fit into your car. Except that, with software, it is something intangible that is not consumed, but can be incorporated time and time again.

In terms of software sustainability, this is precisely what makes it interesting. Because there are sometimes significant economies of scale here. For example, if you have a truly efficient library that becomes widely adopted, the efficiency gains from that single library can be replicated millions of times over on devices running that source code.

The question of where code is executed is also an interesting one here. After all, there are an incredible number of deployment strategies where devices are configured on a large scale. In this context, you set up reproducible systems that can be reused time and again, so you don’t have to do everything manually. And if you can make adjustments there that yield small efficiency gains, then that scales up significantly.

Fascinating! Are there any guidelines on how to design such systems as efficiently as possible?

One thing you should definitely bear in mind is the type of software you’re developing and which part of the software you’re working on. On the one hand, there’s the backend, for processes hidden from users. [Editor’s note: This is the administrative area where developers make changes to websites or software.] Here, the focus is partly on performing calculations more efficiently.

However, efficient software also depends on how people use it. If you don’t have to click around a software for ages to find things, you spend less time using it—and that in itself makes it more efficient in principle. Developers can therefore also design software in such a way that users are guided to be more efficient. After all, efficiency also means making the best possible use of available resources so that they are used as sparingly as possible.

Another interesting aspect is to consider when certain tasks need to be carried out. If, for example, I have a computationally intensive workload—say, if I want to make a climate forecast—then that requires a relatively large amount of computing time. And then I might consider whether I absolutely need these calculations immediately or within the next five hours.

If I need them within the next five hours, I can, for example, shift the computing power slightly if it’s raining at the moment. If the sun is due to shine in the next two hours, I can shift computationally intensive tasks to that time when as much renewable energy as possible is available via the grid and thus save on emissions.

This practice is called carbon-aware load shifting and is technically feasible. Unfortunately, however, I can’t tell you how widespread it is at the moment.

That reminds me of the Grid-Aware websites from the Green Web Foundation. But are there any risks involved in sustainable coding?

Let’s imagine that some genius managed to make the available hardware twice as efficient as it is now. That would, of course, lead to huge savings in energy costs. However, this money is usually invested in new hardware. And so, naturally, there is no further gain in terms of sustainability.

Such rebound effects are often the problem. The miniaturisation of devices alone—the fact that virtually everyone has a smartphone, for instance—means that technology is becoming more widespread and things like language models are being used more frequently. This alone has significantly lowered the barrier to consumption in recent decades.

© Umwelt-Campus.de

Max Westing works in the Department of Environmental Planning and Environmental Engineering at Trier University of Applied Sciences. He has also contributed to the Green Coding project run by the German Informatics Society.

In the past, if we wanted to watch a film, we had to go to the cinema. That involved a consumer choice where you had to pay money, and in return, you got two hours of entertainment. Today, for example, most people have a Netflix subscription, and that’s just part of the consumption decision. At the same time, however, they also have Disney+ or other streaming services. This lowers the barrier to consuming more, as consumption is less conspicuous and content is more easily accessible. Consequently, much more is consumed, which increases our digital carbon footprint.

You mentioned language models a moment ago and also said that a lot of coding is now done using generative AI. What do you think the environmental impact of so-called ‘vibe coding’ will be?

Last year, I co-authored a paper in which we had an AI agent solve some relatively simple problems. We then gave it a few attempts to further optimise the code for energy efficiency. And in a way, that actually worked.

However, it is still difficult to determine what constitutes the quality or ‘standard’ of efficient code. So: what constitutes good code, and how often does one actually come across it in the ‘wild’? After all, there are plenty of developers who do not produce efficient code, either because they do not have the time or because they simply did not need to and therefore never learned how. Consequently, it is difficult to say whether human-written or machine-generated code is more efficient.

What can be said at present is that AI systems that generate code consume a great deal of energy. This is partly due to their training and partly due to the hardware they run on, which is very resource-intensive. However, I am not currently aware of any comprehensive life-cycle analysis that compares whether a developer or an AI agent produces more efficient code.

How can we ensure a green digital future?

Growing e-waste, carbon emissions from AI, data centre water usage—is rampant digitalisation compatible with a healthy planet? Our latest project explores how digital tools and services can be developed with sustainability in mind.

If we assume that developers gain in efficiency when they use AI tools, as they have more time to optimise their code, the question remains as to whether this offsets the costs of training and operating the AI.

That is why there are projects such as ECO:DIGIT, which, if they work well, can provide us with data on this topic.

Thank you for the interview, Max!