At UT, Sjoerd van den Belt (PhD) is working on a project that combines AI, new hardware, and automotive technology. The goal is to make the smart systems in automotives more energy-efficient, so they can do more with less power. In partnership with Toyota, IMEC and other research groups at UT and TU/e, Sjoerd is helping develop the next generation of AI that is lighter on energy but just as capable.
This work is part of a broader push to make computing more sustainable, and it starts with how we build the hardware that AI runs on.
Powerful AI, high energy use
AI is everywhere. It is in our smartphones, security systems, smart speakers, and even in our cars. These systems can recognise faces, understand voice commands, and help vehicles stay in their lane or avoid obstacles.
But behind all this intelligence is a big issue: power consumption. Running AI models, such as those processing images or video, requires a lot of computing. This means more electricity, larger batteries, and more heat. In data centres, this leads to massive energy bills. In cars and mobile devices, it means faster battery drain. That is where Sjoerd’s research comes in. His work focuses on energy-efficient AI, with a special focus on computer vision: AI that allows machines to "see" and understand their surroundings.
Computer vision
Computer vision is a type of AI that processes visual data from a camera. For example, in a self-driving car, computer vision helps the car detect road lines, traffic signs, pedestrians, or other vehicles. It helps the car make sense of the world around it. This type of AI is very powerful, but also very demanding. It often needs to run constantly and respond instantly. That is especially challenging when the device (like a car or drone) runs on a battery. “The kind of computer vision we work on needs to run inside the vehicle, not in a data centre,” says Sjoerd. “So energy use becomes a major factor. You cannot afford to waste power.”
Analogue chips for AI
Most computers today use digital chips. These chips handle data using binary code: 1s and 0s. While they are fast and precise, they can be energy-hungry, especially when running complex AI tasks. Sjoerd’s project takes a different approach. He works with analogue chips: hardware that processes information using continuous electrical signals instead of binary code. These chips are being developed by the Nano Electronics group at the University of Twente.
Analogue computing is not new, but using it for AI is still cutting-edge. It opens the door to faster, simpler, and much more efficient processing. “Analogue chips do not just mimic the brain, but they actually behave more like it,” Sjoerd explains. “Instead of crunching numbers the digital way, they can do it using physics directly. That saves energy.” The project looks at how to combine analogue and digital hardware so that the most demanding parts of the AI model run on energy-saving analogue chips, while the rest remains on traditional systems.
Embedding AI directly into automotives
Instead of sending data to a cloud server or central computer, Sjoerd’s research focuses on embedded systems, that is, running the AI directly on the device, like inside a car. This is especially useful in automotive applications, where a quick reaction time is critical (for example, to detect a cyclist or a stop sign), where privacy is important (data stays inside the car) and where battery life must be preserved (especially in electric vehicles). Working with Toyota, Sjoerd’s team is using real-world automotive use cases to test and improve the system. Although the technology is still in a research phase, it is being designed with real applications in mind.
About the research
The project follows a step-by-step process:
- Model design: identify which parts of an AI system can run on analogue chips
- Hardware integration: combine those parts with existing digital systems
- Prototype: build a working demo using a camera that runs computer vision tasks on the new hardware
This kind of work is known as fundamental research. It is not yet about building products for the market, but about testing ideas, learning what works, and paving the way for future innovation. “What we build now won’t be in cars next year,” Sjoerd says. “But in five or ten years, parts of this could very well be standard in many devices.”
A mix of software and hardware
Sjoerd’s background is a mix of software and hardware. During his bachelor’s and master’s, he worked on facial recognition and hardware acceleration for AI models. He is interested in hardware-software co-design: how tailored hardware can make AI models more efficient. His PhD allows him to combine those interests, exploring not only how AI models behave but how they can be reshaped to work better with new kinds of chips. “It is exciting to work on something so new,” he says. “This is shaping what future AI systems might look like.”
Future
The project will continue for about three and a half years. By the end, the team hopes to have a working demonstration showing how energy-efficient AI can operate in real-world conditions, using live camera data. If successful, this work could influence how AI is built into everything from electric cars to smart cameras to wearable devices. It is a step toward a world where AI can be just as smart, but much lighter on energy.
