Why adopt Green IT, and how? - panel report

Last week, we welcomed four green IT experts to Scaleway - global experts, as France is a pioneer in this field! - to establish how best to move forward on this essential subject. Essential, because digital represents 4% of global emissions (2.5% in France), and according to ADEME and ARCEP, without corrective action by 2050, this proportion will triple.

Hence our invitation to Thomas De Latour, Eco-conception Engineer at ADEME, the French national agency for energy transition; Lise Breteau, Environment & Tech Lawyer, GreenIT.fr, the association that literally defined green IT with its first reference framework, released in 2012; Killian Vermersch, CEO and Cofounder of Golem.ai, experts in frugal AI, since their symbolic AI-based solutions emit 1,000 times less than GPT-3; and Bastien Wirtz, DevOps Engineer at Scaleway, who conducts research on eco-design, and in particular on static websites.

Like our white paper and webinar on the subject, the panel started from the premise that there can be no green IT strategy without acting on the three pillars of data centers, hardware and software. So a holistic approach is key.

But why is green IT such a hot topic right now? A check-in on the regulatory issues was a good place to start.

Regulation: Deadline 01.01.25

In France, as in Europe, January 1, 2025 will see the implementation of two new laws of vital importance for green IT. Firstly, the REEN law, which aims to reduce the environmental footprint of digital technology - and is the first law of its kind in the world - will oblige all French localities with more than 50,000 inhabitants to formalize their responsible digital strategy by early 2025.

The same deadline will apply to CSRD, a European directive that will require all companies with more than 250 employees to provide detailed non-financial reporting, including on scope 3 emissions, which are the most difficult to measure.

"So it's a subject that's going to grow, because of the compliance requirements," explained Breteau. "In addition, there will be obligations on manufacturers and distributors of electronic equipment, or the eco-design of digital services. These are subjects that are increasingly coming to the legislators' table."

For De Latour, this evolution can also be seen within ADEME: "less than a year ago, we only had one person working on the subject of digital sobriety. Today, there are twelve of us. It's our job to provide both the public and private sectors with the tools they need to measure our impact and respond appropriately to existing legislation and standards, which are set to become even simpler."

On the private sector side, Golem.ai can also feel the tide turning. "Five years ago, in calls for tender, the ["sustainability"] criterion was there. Now, and increasingly so, it has become a dealbreaker," says Vermersch.

According to Scaleway's Wirtz, things are starting to change on the developer side too. "Quite a few developers have an approach that is compatible with eco-design. When you look at the eco-design guidelines for websites, it's really just about doing things right. We can analyze what each function in a code is going to consume in CPU," with tools like SonarQube, which lets you test the efficiency of your code, and whose green IT plugin, ecoCode, lets you go even further in terms of energy analysis.

That said, measurement indices at all levels remain one of the main stumbling blocks to be resolved. "Do we all have the same analysis grids? And what do we measure? This is an absolutely essential issue," asserted Breteau. GreenIT.fr recommends the PEF method, or Product Environmental Footprint. This grid, developed by the European Commission, proposes sixteen measurement criteria for digital impact, "including water, including resource depletion", she added. "Impact is not just CO2. 50% of the digital footprint is linked to the extraction of non-renewable resources."

Data centers: How to optimize your cloud impact

If data centers are a major contributor to the impact of digital technology - they account for 1% of the 4% of emissions mentioned above - cloud customers have every power to limit this impact (provided they choose a responsible provider like Scaleway, of course!) And for those who do, the benefits can be both economic and ecological. Such is the case with Golem.ai.

Firstly, as Vermersch explained, by using frugal technologies themselves. As Golem.ai mainly uses symbolic AI, which is far less resource-hungry than generative AI, "per month, we pay our cloud bill at around €10,000. Which is close to nothing. Often, an AI start-up that needs to train models will be at €50,000 in its early days, and that bill will quickly go up."

Secondly, as we explained elsewhere, Golem.ai managed to halve the number of cloud machines needed to process millions of emails a month thanks to auto-scaling of its Kubernetes infrastructure, essentially by turning off unnecessary clusters. "We've halved our costs, while tripling our business volume at the same time. So apart from software optimizations, we've divided our resource consumption by five or six, roughly," explained Vermersch.

And let's not forget the other advantage of the public cloud in terms of sustainability: pooling cloud resources means that Golem.ai will be able to access new clusters or servers quickly in the event of load peaks, because the company uses the same hardware as two to ten other companies… and so its use of auto-scaling will free up resources for other Scaleway customers.

Hardware: Repairability and slow tech

With hardware accounting for the majority of digital impact, extending the life of devices remains one of the most effective ways of reducing emissions. That's why ADEME has developed the Repairability Index, another French innovation, to help consumers make better choices before they buy. The index clearly indicates "what the designer of the product has planned so that you can keep it for as long as possible" said De Latour. "Even though it's something we've implemented nationally, the big manufacturers have started to rethink their design [globally]. It's an incentive for eco-design."

It's for similar reasons that GreenIT.fr advocates for slow tech, or low tech, or the principle of "running [new] services on old hardware," explained Breteau. "The area that's going to have the biggest environmental bill is devices. So it's important to have digital services that are accessible. We shouldn't necessarily need the latest smartphone to access a travel website or online banking."

"We can make weather forecasts that are ultra-precise, thanks in particular to supercomputers," she continued. "And that can be very interesting, for farmers for example. Why not distribute this information via simple text messages on old [Nokia] 3310s? That's precisely what the Weatherforce project has done, for example. Combining powerful high-tech capabilities with very basic standard techniques avoids the need to renew terminals, and therefore waste."

So, developers: ready to imagine your next SMS-delivered news app?

Software: Websites and eco-design

Data centers and hardware therefore account for the majority of digital emissions. But all these machines run on code, which we don't always think about optimizing in terms of energy consumption. And yet we should. Websites today, for example, weigh 200% more than they did ten years ago. Why is this? And how is eco-design a solution?

"The DOOM designers built those games on machines that were very basic, technically speaking. That's quite a feat. Today, we don't need those skills anymore," said Wirtz.

So it's up to developers to try to minimize the energy consumption of their work, according to Wirtz, for example by avoiding the use of heavy images, or autoplay videos, on their websites. One great example of eco-design are static websites, an alternative to solutions like Wordpress, which Wirtz explored in our white paper, and here.

"If we take the metaphor of a museum, WordPress will paint you the Mona Lisa every time you come. Whereas a static site will paint the painting once, and then afterwards, you'll keep consulting [the same painting]. Even if the data changes, the painting will be redone on demand, not in real time every time." This reduces the amount of data required each time a website is consulted, and therefore its energy consumption.

"That said, a static site is just a tool, it doesn't solve anything" [alone], Wirtz clarified. Scaleway.com, for example, is designed with a static site generator, but its score on EcoIndex [GreenIT.fr's tool that measures the impact of websites] is relatively low, not least because it contains "five videos that are over 6MB", according to Wirtz. "So our site is almost perfect without those videos. That's really not a criticism of the team! We may have a very light foundation, but we also need to build lightly on it, ideally."

Frugal AI: dream or reality?

For a little over a year now, everyone has been talking about generative AI, to the point where its potential impact on the planet can be overlooked. ADEME is one of the organizations now looking at how to limit this impact, with the aim of imposing "frugal AI", particularly in the public sector.

"Today, AI is everywhere," asserted De Latour. "We have to try to make it frugal. It sometimes feels like using a kettle to heat water to just 30 degrees. It's not adapted to our needs at all. We also have to think about accessibility. What can give me access to AI and, we hope, frugal AI to everyone?"

This can be achieved through a slow tech approach, Vermersch pointed out, with the example of "llama.cpp, an open source application that has managed to translate Meta's Llama model into C/C++, which means it can run on older machines (e.g. a two- to three-year-old Mac). This goes a long way. Today, there are more and more open source alternatives."

We can even imagine, as shown by the example of Mistral AI at ai-PULSE last November, that AI engineers have every interest in making their models as small, and therefore as energy-efficient, as possible. What's more, there are almost 500,000 different models on Hugging Face; we can only hope that the majority are not monstrously resource-hungry…

"The real question is how do we train buyers to make the right choices?" asked Vermersch. "We can also use an AI that will cost a fraction of the price (of ChatGPT), that will also structure the info; and it works. We've known how to do that for a long time.** 90% of what's done with [OpenIA's GPT models] has limited value. A lot of people use them for fun, or for small personal projects**. That's all well and good; but for anything that involves workflow automation, information extraction, document processing, etc. [symbolic AI is more than enough]."

Which brings us back to the fundamental question of digital sobriety: do I really need this latest innovation? Since GPUs for AI can consume between 3 and 5 times more energy than equivalent CPUs, shouldn't we reserve this firepower for the most useful applications? And yet, at the end of 2023, one of the most popular models on the GPT Store was a plugin for ChatGPT that allows you to fold your laundry properly (source, and the app in question).

More than ever, "training buyers to make the right choices" will be the key issue for all responsible digital in the not-too-distant future. So stay tuned!