How do we make technologies more efficient and less energy consuming?
In the age of resource awareness, ensuring the energy efficiency of digital technologies shaping our world is critical. Associate Professors Pınar Tözün and Zoi Kaoudi from IT University of Copenhagen tackle the issue from a hardware and a software perspective. Catch their session at this year’s Digital Tech Summit, 31 October.
Pinar TozunZoi KaoudiComputer Science DepartmentEventsResearchClimate IT
Written 10 October, 2024 08:11 by Theis Duelund Jensen
According to a recent report from the International Energy Agency (IEA), data centres consumed an estimated 460 terawatt-hours (TWh) in 2022 and could reach a total electricity consumption of more than 1,000 TWh in 2026. To put that astronomical figure in perspective, it is roughly equivalent to the annual electricity consumption of Japan.
Powering the large-scale technologies that act as cornerstones in the modern, digital economy has become a critical issue in light of global conflicts and environmental concerns. That is why the field of performance optimisation has become a hot topic in the industry as well as in academia and research. Associate Professors Pınar Tözün and Zoi Kaoudi, both from IT University of Copenhagen’s Computer Science Department, tackle different sides of the issue.
The two will present their ongoing projects and findings at this year’s Digital Tech Summit in Copenhagen on 31 October.
Making the most of existing resources
Pınar Tözün focuses on optimising hardware utilisation for data-intensive applications like artificial intelligence and large databases. "Hardware quality is constantly improving," she explains, "but software doesn't automatically adapt. To leverage this improvement, we need manual software adjustments for better efficiency."
Highlighting the enormous carbon footprint of training large language models – some consuming more energy than a person's lifetime usage – Pınar Tözün emphasizes the need for smarter software. Her research strives to achieve the same computational results with fewer hardware resources. "The key is maximizing performance from existing hardware," she says.
Her work starts with extensive performance analyses, essentially building a profile of the hardware's strengths and weaknesses. This information allows them to develop techniques that enhance performance at the software level. "We may introduce additional software layers or adjust existing ones," she explains. "This allows devices to handle multiple applications without sacrificing performance. Imagine running two large deep learning models on a single device, eliminating the need for a second one."
Standardisation, however, presents a challenge. While GPUs from the same vendor tend to behave similarly across generations, different vendors require different software adjustments. The researcher likens the process to a medical check-up: "We perform in-depth analysis to understand the hardware and discover its optimal use."
Optimising data processing
In her research, Zoi Kaoudi explores the use of AI for database optimisation, particularly search queries:
"Query optimisation is a cornerstone of any data management system," says Zoi Kaoudi. "Whenever you interact with a computer, whether it's executing a command or retrieving information, you're essentially querying the system. Efficient query optimisation is crucial for minimizing runtime, processing power, and ultimately, energy consumption."
The researcher’s work leverages the power of artificial intelligence and machine learning to enhance query optimisation. "These technologies have proven to be invaluable tools in automating tasks and improving efficiency," says Zoi Kaoudi who, in addition to her position at IT University of Copenhagen, is co-founder of Scalytics, a startup specialising in software solutions that optimise data processing efficiency and reduce energy consumption.
"We haven't fully realised the potential of AI and machine learning in query optimisation. In the project I am currently working on, we delve into the core aspects that influence efficient optimisation: the relative order of query plans and the underlying ranking algorithms. By analysing the connections between queries and query plans, developing learning-to-rank models, and creating new optimisation algorithms, we aim to pioneer a new research paradigm," says Zoi Kaoudi.
Find more information about Pınar Tözün and Zoi Kaoudi's talk at Digital Tech Summit here.
Theis Duelund Jensen, Press Officer, phone +45 2555 0447, email thej@itu.dk