On the way to the AI revolution
Artificial intelligence is seen as the engine of the future. However, the enormous computing resources that large language and learning models require are increasingly raising questions about sustainability. In conversation with ETH computer science professor Ana Klimovic, it becomes clear that increases in efficiency are possible and crucial in order to bring the growing hunger for energy of AI under control.
The debate about the power consumption of AI systems is not just a political issue. Data centres and highly scaled hardware consume enormous amounts of energy and the constant increase in the size of models is further exacerbating this trend. “We can’t scale indefinitely,” explains Klimovic, “so research into more sustainable solutions is essential.”
Economical model architectures
One approach is the introduction of sparsity (density reduction) in neural networks. Models only activate relevant parts of their system, whereas classic approaches always utilise the entire network. “Mixture-of-experts models follow this logic. They distribute queries specifically to specialised modules. This saves energy without sacrificing quality.
GPUs are valuable, but often unused
Klimovic sees a central problem in the low utilisation of GPUs, even though they consume an enormous amount of power. Bottlenecks occur in particular during data pre-processing and communication between several GPUs. Computing utilisation is often below 50 percent. New software solutions are needed to prevent valuable resources from lying idle.
Efficiency through intelligent frameworks
Your research group develops systems that focus on automation and optimisation.
Sailor is a platform that automatically parallelises training jobs via GPUs, thereby increasing GPU efficiency.
Modyn and Mixtera are systems for smarter data selection that train models faster and with less data.
DeltaZip is a platform that efficiently manages fine-tuned model variants. It compresses differences between models (“deltas”), which reduces loading times and makes inference faster and more resource-efficient.
Sustainability in training and inference
Efficiency gains play a key role not only in training, but also in the application, known as inference. In view of the billions of daily interactions with chatbots, the conservation of energy and hardware resources is becoming a globally urgent task.
Academic freedom and open science
Klimovic emphasises the importance of academic research. Less driven by economic constraints, it can pursue long-term innovations. She emphasises the role of the Swiss AI initiative, which was launched in 2023 and is based on the CSCS’s almost climate-neutral Alps supercomputer. With over 10 million GPU hours and CHF 20 million in funding, it is the world’s largest open science and open source initiative for basic AI models.
The AI revolution will only be sustainable if efficiency becomes the guiding principle. In algorithms, hardware and system architectures. Projects such as Sailor, Modyn and DeltaZip show concrete ways in which enormous energy savings can be combined with technical excellence. For Klimovic, one thing is certain: “In the future, high-quality AI will not only mean intelligence, but also resource conservation.”