Green by design: What is the carbon impact of artificial intelligence on the environment?
When we talk about artificial intelligence, we don't think about the environmental impact of algorithmic models. The focus is only on process and cost optimisation. However, AI has a real impact on the environment. According to one study, algorithms emit as much CO2 as a car during their lifetime.
How do algorithms generate CO2? And how can we reduce their impact on the environment?
Data storage : Highly polluting data centres
Major advances in new technologies have made it possible to store large volumes of data. To accompany this trend, Cloud providers such as OVH or Azure are offering increasingly cheaper storage and computing costs.
Companies can now generate large amounts of data with the computing power to accelerate the prototyping of predictive models.
The cost of storing data and the power of algorithms have thus enabled the democratisation of AI and Machine Learning. Indeed, according to a survey conducted by CCS Insight, around 80% of companies surveyed have already used AI or are experimenting with AI.
However, storing data requires significant energy resources to carry out the processing of these calculations. What is the impact of this computational processing on the environment? What can we expect in the future?
Artificial intelligence: a carbon footprint that is very much present
Machine Learning algorithms consume a significant amount of energy to generate the computational processes. Indeed, the massive storage of data has generated a significant craze around AI. The ease of access to data storage has led to the development of increasingly complex algorithms, which means that more and more computation is required and thus a considerable increase in CO2 emissions. For example, to train an algorithm on natural language, the equivalent of 355 years of computation were required on a single processor. This represents in CO2 emissions, the 700,000 km journey of a new car.
Moreover, the environmental weight of data centres is constantly evolving until it reaches more than 86% of the total digital footprint by 2040. (source: LesEchos)
The reality of AI consumption in companies
Some companies are aware of the greenhouse gas emissions generated by their algorithms, particularly in terms of the ethical but also financial aspects. Indeed, the cost of storage remains high when large volumes of data and calculation processing are generated.
In addition, carbon emissions can be more significant depending on the type of energy used to power the data centres. Moreover, servers are sometimes installed in countries using a carbon-based energy resource (coal, oil, gas). Some AI companies nevertheless try to source their energy from suppliers who use only renewable energy, which emits the least carbon.
Reducing the carbon footprint of algorithms thus involves the choice of cloud server, in particular the location of data centres.
Unfortunately, some cloud providers do not communicate any information on energy consumption and greenhouse gas emissions. This makes it impossible to measure the impact in terms of CO2 emissions without knowing how much energy is used to extract the data.
Consequently, it is difficult to measure the impact generated by algorithms on the environment.
Conflicting issues for companies
Optimizing operational efficiency and customer experiences has become a major issue for many companies, especially those with an environmental focus. The environmental issue is the one that attracts the most interest since it has a significant impact on financial costs.
The paradox encountered by these companies is the will to implement use cases to optimize carbon emissions but using an AI that itself generates a certain carbon footprint. As a result, the positive effects generated by the use case are cancelled out by the CO2 generated by the computational processing of the algos.
In the future, companies should have a think about the implementation of their use case, including the choice of cloud providers but also the choice of the AI platform used.
Towards a more sustainable use of AI
A few years ago, a consortium of companies set up an open source platform, called Code Carbon, to help companies measure the impact of their code and algorithms, and thus measure the carbon footprint of its work.
Through the platform, it is thus possible to measure the CO2 emissions of one's work and to have an equivalence with simple indicators, notably the number of kilometres travelled by a car.
This consortium thus has the will to raise the awareness of companies to the use of AI, in order to engage them towards an ethical AI.
However, it is not enough to use tools to measure the impact of one's work, a reflection is needed on the tools used to minimise the impact of one's code. Indeed, not all AI platforms used to implement use cases adopt an ethical approach. Some of these platforms generate a significant amount of data storage, particularly due to the computational processing performed.
Consequently, the choice of AI platform used to implement these use cases will have a significant impact on data storage, and therefore on its carbon footprint.
The papAI platform allows to measure the equivalence in terms of CO2 emitted and renewable energy needed to perform intermediate processing.
This approach helps to make developers aware of the impact of their work on the environment, through the CO2 equivalence of processing a calculation. For example, the data workflow generated to measure the prediction of customer churn represents the energy consumption of a dwelling over a day. Through this, it is possible to deduce the cost of each calculation process and thus think of a method to reduce costs.
The use of the papAI platform thus allows for a more responsible approach than some platforms. Instead of storing all the intermediate data, which generates high storage costs and carbon emissions, papAI stores only part of the intermediate data. Indeed, intermediate data does not need to be systematically stored. The papAI platform will judge the importance of these processes by the computation time that was required on previous datasets. If a computational process has taken 24 hours, it will be more favourable to store this data. This approach allows an intelligent choice of data and processing to be stored. Through the papAI platform, it is possible to build your own processing scenario to group together lots of small processes and transmit the data from one step to another without the need to be stored.
Through this approach, papAI allows to simplify the processing flow, to optimize storage and to reduce the computation time which generates a consequent cost. The papAI platform thus adopts a Green by Design approach in accordance with the environment.