The chips of the future

‘The computers we have today are not designed for AI,’ says Beatriz Noheda, professor and director of the Groningen Cognitive Systems and Materials Center (CogniGron). As a result, they use an unnecessary amount of energy, and we are also reaching the limits of our current technology. That is why CogniGron is working on new materials that can mimic the human brain. CogniGron professor Tamalika Banerjee will soon be developing chips for energy-efficient computers with her start-up IMChipN.

“Imagine walking into a room full of people, looking for one specific person,” says Beatriz Noheda. For a human being, that task is a piece of cake, but a computer, even modern AI-based Beatriz Noheda facial recognition, is not very good at this at all. The computer will study all the pixels one by one, which takes an enormous amount of energy. Whereas we humans can quickly prioritize (only look at faces) and check multiple things at once (black hair or blonde, glasses or no glasses). Noheda: “To recognize a face, a computer needs about 20,000 watts, while a human being only needs an estimated 20 watts.” And that's despite the fact that much of modern AI is inspired by the human brain – such as the popular Machine Learning technique, in which a computer learns to recognize images by ‘practicing’ with a lot of training material. The problem is that we run these AI algorithms on digital computers that are nothing like the brain. That's precisely what makes AI such an energy guzzler.

Also between zero and one

The human brain works structurally differently from a computer: it does not store information as zeros and ones, but as everything in between. And new information is stored, as it were, in the hardware itself: in the connections between neurons. For example, when a child learns the connection between the sound ‘dad’ and the face of their father, they also develop physically strong connections between neurons in the brain. Noheda: ‘The brain does not distinguish between hardware and software, or between memory and processing. Everything is integrated.’ In addition, the connections in the brain can be weak or strong, or something in between – very different from the transistors on a computer chip, which are only ‘on’ or ‘off’ and store information as zeros and ones.

IMChipN

According to Tamalika Banerjee, current computers are limited by three problems: (1) memory and processing are not in the same place, (2) it can take one to two seconds to send data back and forth to the cloud, and (3) after years of innovation, we are now reaching the limits of what is possible with current materials. That is why she wants to offer new hardware with IMChipN that is easy to manufacture, durable, and can be combined with existing technologies. The company will develop a chiplet, a small custom chip for a specific task.

Memristors

Noheda explains what a difference this makes: “To mimic a single neuron, you need a lot of transistors. We work with new systems called memristors: hardware that gradually evolves with use, depending on how much current has passed through it in the past.” Just like a child forming stronger connections in the brain. Noheda himself works with crystals in which a network of lines can conduct electricity. The more electricity that has passed through them, the better they conduct. ‘That's how we learn!’ says Noheda. ‘And we can grow this crystal on a nanoscale in our lab.’

Without delay

“If you open up your laptop,” says Tamalika Banerjee, “you'll see that the processor and memory are in two different places.”

In addition, we send a lot of data to the cloud, and it takes one to two seconds to retrieve that data. And here's another fact: a single data center (the cloud relies on many such data centers) uses about as much power as 80,000 households. Keeping memory and processing close together will therefore be more energy efficient. “What's more, the speed of that data exchange also limits the speed of the computer as a whole,” says Banerjee. “Think of medical applications, self-driving cars, but also the gaming industry, for example: these require instant decisions, without any delay on the line.”

Startup in the North

IMChipN, Banerjee's startup in the making, will manufacture in-memory chips using a combination of old and new materials, with memory and processing taking place in the same place.

The N in the name stands for North, because Banerjee wants to keep the company here in the region. Her PhD students have already been working in the lab on developing the necessary new materials and devices: “We have the NanoLab here and our own CogniGron labs, where we have already done some promising work.”

Developing new materials and devices

Tamalika Banerjee has worked with silicon chips for many years. “But current materials limit the possibilities of devices,” she explains. In Banerjee's lab, PhD students are working with complex oxides.

“These materials have many more possibilities than silicon,” says Banerjee. “You can process them to become a semiconductor or a superconductor. You can make them ferroelectric or change their magnetic order.” These are all physical properties that determine the behavior of the material, and by using them cleverly, you can create a memristic device. Banerjee explains: ‘We can process the material in such a way that it behaves like a network of neurons and the connections between them, and that, like those neurons, it can take on values between zero and one. We have shown that it also works very well on a nanoscale. In fact, it is precisely on a small scale that we have achieved spectacular results.’

Text: Charlotte Vlek