The demand for faster, more powerful computer systems continues to increase, partly due to innovations in AI and data analysis. At the same time, sustainability is becoming increasingly important, and the energy consumption of the sector is a major burden on the environment.

For example, a single search on ChatGPT costs significantly more energy than a search on Google. Training AI models also costs a huge amount of energy. ‘That's just not right, it’s not sustainable. We have to do something about this’, says Andy Pimentel, professor and chair of Parallel Computing Systems at the Informatics Institute (IvI) of the University of Amsterdam.

Pimentel focuses his research on computer systems that are incorporated in products, such as mobile phones, televisions, and ASML's advanced chip machines. His research group uses technological models to analyze and improve specific behavior of computer systems. Pimentel explains: 'We are mainly interested in how the computer does what it’s supposed to do. For example, we try to increase the speed or reduce energy consumption.'

Energy labels

So, Pimentel's research group uses technology to make computer systems more sustainable, among other things. An example of this is the Energy Labels-project. For instance, when you have a Zoom conversation, you use an entire digital chain of software and servers, all of which consume energy. The researchers first want to map out how much energy is consumed in this chain using models. Then they want to be able to influence this chain.

Pimentel: 'What we want to achieve is a kind of energy label system for the digital chain. For example, it could indicate: 'Now the label is C, but if you switch to another service, the label will be A'. This way, we can encourage the user to perhaps opt for a more energy-efficient solution.’ The researchers also collaborate with the economics and law faculties of the University of Amsterdam.

Making AI possible

Optimizing computer systems is also essential for the AI revolution. Pimentel: ‘AI has been able to gain its current popularity due to three factors: the AI algorithms have improved, there is now a lot of data available to train the models, and the computer systems have become fast enough so that large models can be trained. The latter is often forgotten, but it is crucial.’

Pimentel’s research group also conducts a lot of AI-related research itself. For example, they're working on the possibility of executing large, complex AI algorithms on small computer systems. Normally, these algorithms are executed in the Cloud, where a lot of computing power is available. However, this can be problematic if there is no stable internet connection.

Pimentel: ‘If you're in a self-driving car, you don't want the AI, which analyses whether a pedestrian is crossing, to lose its network connection. That's why we want to implement these large neural networks in AI algorithms as close to the user as possible. Our challenge is how we achieve this.’ The researchers do this, for example, by chopping up the large neural network into several pieces and distribute these pieces across mulitple small computer systems close to the user, such as a mobile phone. In this way, the systems together deliver the final result.

Modelling as a common thread

Improving one aspect of a computer system often comes at the expense of another. For example, energy consumption often increases when you want to make the system faster. Pimentel’s research group describes these aspects in models, so that they can then optimise the system. ‘We try to explicitly map out the trade-offs for someone who wants to design such a system.’

Technological models form a common thread in their research. But the number of possibilities for designing a system is often greater than the number of stars in the universe. To achieve good results, the researchers use all kinds of smart algorithms.

Nature as a source of inspiration

They are inspired by algorithms that imitate nature. Pimentel: ‘A good example of this are genetic algorithms. With these, we describe a solution in the form of a chromosome. Each part of that chromosome represents a choice that you make, in which you can make small changes.’

Another example is “ant colony” optimization, which mimics the behavior of ants that search and leave trails. These algorithms help researchers find solutions for optimizing computer systems. In this way, they work towards a sustainable future with advanced technology.

Do you want to know more about research at the Faculty of Science? Take a look at our research page.