Science for Society | The AI chip of the future
Scientists work day by day on solutions to a variety of problems. From new drugs to smart farming techniques, our research helps society move forward. But this is not always immediately visible. In the Science for Society series, we use stories about impactful innovations, products and ideas to show that science works!
Text: Thomas Vos, Corporate Communications UG
A thousand times greater performance, two hundred times more energy efficient and a hundred times more data transfer speed: the startup IMChip is working on the AI chips of the future. Among other reasons, their work is crucial in times when concerns about the vast amounts of energy used by AI are rising.
Ruben Hamming-Green is co-founder of IMChip alongside Professor Tamalika Banerjee and Azminul Jaman. They all have backgrounds within CogniGron, a cross-disciplinary research institute that takes expertise in materials science, mathematics, AI and chip design to research computing inspired by the human brain. Hamming-Green: ‘We built IMChip from scratch, using Professor Banerjee’s research and ordering the right equipment. At the moment, we are still researching and developing in the cleanroom here at the Feringa Building. Our first round of funding came through in June, from Future Tech Ventures, of which RUG Ventures is a limited partner,’ Hamming-Green says.
Synapses
What exactly is IMChip trying to do? To explain that, you first need a basic understanding of the current situation regarding AI and computer technology. Hamming-Green: ‘Current computer technology is built all around transistors, which are, simply put, on and off switches, the zeroes and ones that often come to mind when thinking about computers. Very useful if you want something to act in a controlled and defined way. However, it is really different from how we act as humans. We don’t exactly know how our biological processing works. AI, however, is trying to mimic exactly this. It is trying to work in the same way as our brains do, where synapses form the connections with neurons. As we use them, synapses get stronger or weaker. This strengthening and weakening is how information is encoded in our brains. It is not a zero or a one, it is everything in between.’
The technology: memristors
Most AI is still using transistors, however. And that is exactly what Hamming-Green and his colleagues want to change: ‘We are building chips that use memristors, which basically act like the synapses in our brains. When you pass an electrical signal through it, it can be modified to have different levels of conductivity. You can give each element a programming sequence that will make it either very conductive, medium conductive or not conductive at all. A memristor can do this in a continuous manner.’
Two million homes
With the AI boom happening, this technology is needed, Hamming-Green explains: ‘Deep neural networks such as ChatGPT need lots of transistors to do tasks that transistors were not built for originally. Furthermore, there is the issue of processing and memory. Within our smartphones, these are separated, but that is not the case for our brains. There, they are the same thing. With the current technology, this means that you constantly have to switch between processing and memory in AI applications. This not only takes a lot of time, but also large amounts of energy. And that is a very pressing issue. A company like OpenAI uses the energy of about two million homes to keep their systems running. Memristors take all these different processes and reduce them to single devices. Basically, they are chips designed specifically for AI applications. And they are much faster. You can supply a signal to it, and it will immediately read it out.’
Integrating the technology
IMChip is now facing an important challenge: that of integrating the technology. Hamming-Green: ‘Our whole chip industry at the moment is built around silicon. Current factories are optimized for this. We don’t want to fully move away from that, we will still need transistors, and the current production technology is very robust and mature. However, silicon just doesn’t behave like memristors do. We have to use more advanced materials and find a way to integrate them with silicon, in a way that it can also be integrated with foundry technology. Only then will it be of interest to larger companies. We need to prove to them that we have the integration technology down.’
Memristors are the future
This will still take time, Hamming-Green emphasizes: ‘Our goal is to have a prototype in the next two years, so we can prove that the technology works. After this, we want to develop the first actual product that we can sell to pilot customers. I think we have a good story to sell, though. Other companies are still working with silicon. Their chips are more energy-efficient than current chips, but not as powerful. We see this memristor technology as the future. It just hasn’t been industrialized yet.’
Read more
Do you want to know more about Science for Society? Check out the overview page for the previous editions.