Home

A Human Brain is More Energy Efficient Than AI Chips

Huai Wang has mapped the development of AI tools in power electronics since the 1990’s.

A Human Brain is More Energy Efficient Than AI Chips

Huai Wang has mapped the development of AI tools in power electronics since the 1990’s.

By Niels Landbo Krogh, AAU Communication & Public Affairs
Photo: Huai Wang

The potentials of AI technology became known to larger society in 2022, but Professor Huai Wang has used AI tools in his field of power electronics for a long time. His keynote at AI for Energy – Energy for AI will focus on two subjects: His use of AI as a tool in energy engineering and the larger battle to harness the energy conversion and consumption of AI.

Huai Wang is author of one of the first scientific papers reviewing the AI use in power electronics historically, and it has become popular worldwide: “An Overview of Artificial Intelligence Applications for Power Electronics”.

- Energy planning is one really interesting and potent area we use AI in right now. It’s the hunt to minimize the cost of both energy and equipment maintenance, Wang says.

Along with his colleagues the AAU ENERGY researcher is constantly tuning design parameters on his console or in the labs to reach new desired targets. And the nature of these goals has changed, since he started to see AI use in power electronics.

Energy planning is one really interesting and potent area we use AI in right now. It’s the hunt to minimize the cost of both energy and equipment maintenance.

AI since the 90’ies

How has AI use evolved in energy engineering?

- People started to explore AI in 1990's for power electronics application, which is my main research field, but back then it was really the early tools, such as expert systems and fuzzy logic. The last 10-15 years we have started to use more machine learning. It was only until 2018 that we started seeing broad use of AI technology in power electronics articles on universities worldwide.

Today the vast majority of Huai Wangs hours are also used with models on a screen instead of in a lab. Huai and his team work a lot with digital models in power electronics. This means they build, test and optimize virtual energy systems, large and small – From the simplest such as a small hardware unit with power supply to the most complex such as a whole city with power consumption and production. Each system has its own flows of energy, waste and errors.

Mostly Avoiding Big Data

Working with AI in engineering could be very different than AI in other fields, Huai says.

- From 2015 to 2019 we became more familiar with machine learning. The grand challenge of using AI in engineering back then as well as today is that we can't afford even small errors, cause they can lead to failure. This means fewer use it and narrower, so there is a gap between the literature on AI and how the industry appliance is.

To outsiders maybe only familiar with generative AI like ChatGPT, he explains how AI in power electronics is a different paradigm.

- We don’t do a lot of “big data”. It’s not the answer in this kind of engineering, since the computational resources are limited - there's often not room for using larger chips in most power electronics, or it will make the devices too expensive, Huai explains.

- So 80 percent of the team’s effort lies in getting the right data based on in-depth engineering insights instead of getting all of the data.

At this year’s Research Day, AI for Energy – Energy for AI, Huai Wang’s keynote will elaborate on AI in the field of power electronics, and how power electronic technologies will help power AI chips in a more efficient and reliable way, and he will also unfold part of the process of his work with Danfoss Drives on predictive maintenance.

Potential for Mix in AI Technologies

Large Language Models (LLM) and Multi Modal Learning (MML) are AI appliances that use existing databases to do the training and algorithms. But it’s different in engineering to a large extent, because the field does not have databases in the same way.

- It will be really interesting when at one-point we can use specialized AI databases/systems like GPT's for engineering together with LLMs and MML and we could learn faster, Huai Wang ponders.

For now Professor Wang’s team goes with a “light and right” data focus in most of their AI work. An example of this is a research project that ran from 2020-2024 where Huai with a grant from Villum Synergy worked along with Dept.  of Computer Science at AAU on “Light-AI” technology.

... our human brain is very energy efficient actually. And if we don’t get better at controlling, GPU loads for example, demand for energy will be too high.

- Chips for AI are not that smart and they use way too much energy right now, Huai says.

- As a comparison our human brain is very energy efficient actually. And if we don’t get better at controlling GPU loads for example demand for energy will be too high.