In a recent Universiteit van Nederland podcast, Prof. Dr. Ir. Wilfred van der Wiel explores whether the brain can help solve AI’s energy challenges. AI grows more powerful (and power-hungry), and Wilfred and his team at UT are exploring a very different kind of hardware. Not bigger, faster servers, but brain-inspired chips that do more with less energy. “One AI prompt can use as much energy as charging your phone for ten minutes,” Wilfred explains in the podcast. “The data centre is out of sight, but the energy bill is not.”
Smarter chips that act more like neurons
At the NanoElectronics group in Twente, researchers are building so-called neuromorphic chips. These devices are analogue, trainable, and behave more like networks of neurons than traditional processors. “Their behaviour is not fixed at design time,” Wilfred explains. “You can tune them after production, just like you would train a neural network. By leveraging reconfigurable nonlinear processing units, we can perform complex operations more efficiently, reducing both the number of steps and the need to repeatedly fetch data from memory. The physics does the computing for you.”
One application for this approach is speech recognition. Wilfred’s team recently filed a patent for a system that processes raw audio directly on their neuromorphic chip. This makes it possible to perform high-quality speech recognition with much smaller and simpler digital models. “We skip a lot of the heavy signal processing up front. No need to convert to digital or to the frequency domain,” he says. “That saves a lot of energy and still gives good results.”
What happens when hardware starts to learn?
The idea of a chip that can adapt to its environment sounds powerful, and it is. But it also raises important ethical questions. “We are developing physical systems that can adapt and learn, essentially bringing artificial intelligence into the hardware itself”, Wilfred notes. He and his colleagues are working closely with ethicists and philosophers. “If these systems become more autonomous, we need to ask hard questions about control, responsibility and unintended consequences,” he says.
Wilfred does not see neuromorphic chips as a replacement for digital computing. “We will always need precise, robust systems for a lot of tasks,” he says. Our adaptive chips could make a real difference for data-intensive jobs where energy use matters, like edge devices in phones, cars, or medical implants. “They are lightweight, fast, and do not rely on a constant internet connection. That is what makes them ideal for real-world environments,” he adds.
As AI hardware evolves, so do the societal impacts. Wilfred hopes more researchers will take this seriously from the start. “We do not want to build a powerful technology and only later ask whether we should have done it differently,” he says. “That conversation has to happen now, as the field is still taking shape.”
