New AI chip for in-memory computing

Computing-intensive applications such as AI or big data are based on highly efficient microelectronics, which will be made more powerful and at the same time more energy-efficient in the future through ferroelectric field-effect transistors.

Digitalization, big data, or artificial intelligence (AI) place high demands on hardware, which is finding it increasingly difficult to meet the ever-increasing demands for more computing power and higher energy efficiency. Energy-saving concepts such as neuromorphic computing should help remedy the situation. It attempts to replicate the most efficient and flexible memory in the world—the brain. Chips based on non-volatile memory devices such as field-effect transistors made of ferroelectric materials (FeFET) are particularly promising.

They change their polarization when an electric field is applied and retain this state after the voltage is switched off. The multiply-accumulate operations (MAC) typical of deep learning algorithms can therefore calculate and store the weights in the same chip (in-memory computing). That eliminates time-consuming data transfers between processor and memory – known as von Neumann bottlenecks.

As the only non-volatile memory concept, ferroelectric memories are also operated purely electrostatically and are therefore particularly energy efficient, since only the recharging currents of the capacitors are required to write data.

FeFETs on the rise

The promising properties of this fascinating storage technology make it the subject of a host of research projects worldwide with, in part, very different approaches.

Researchers at the University of Pennsylvania School of Engineering and Applied Science, for example, developed an FeFET design with record performance for data processing and storage. The super-thin sandwich transistor layers a 0.7 nm thick semiconductor made of molybdenum disulfide (MoS2) onto the 20-nm-thick ferroelectric material aluminum scandium nitride (AlScN). The unique ferroelectric properties ensure reliable data storage even with the smallest form factors.

FMC and the Fraunhofer IPMS, on the other hand, are focusing on ferroelectric hafnium oxide (HfO2), currently the standard gate dielectric in ferroelectric memories. That makes them CMOS-compatible, lead-free and scalable down to very small technology nodes.

Energy-saving AI chip from Munich

Together with Fraunhofer IPMS and Bosch, which has a stake in FMC, the TUM professor Hussam Amrouch recently presented an AI chip with ferroelectric field-effect transistors (FeFET) based on hafnium oxide. At 885 TOPS/W (tera operations per second per watt), it is said to be twice as powerful as comparable in-memory computing approaches (e.g. Samsung’s MRAM). The FeFETs are manufactured by the U.S. company GlobalFoundries in a 28 nm process in Dresden.

In an interview with authors of the Nature journal, Professor Amrouch stated that his chip achieved an accuracy of 96.6 percent in handwriting recognition and 91.5 percent in image classification without additional training.

The AI chip is to be used wherever data needs to be processed directly where it is generated, such as generative AI, deep learning algorithms, recognizing objects in space, or for robotic applications. However, the scientist from the Munich Institute of Robotics and Machine Intelligence (MIRMI) thinks that it will be three to five years, at the soonest, before the first in-memory chips suitable for real-world applications become available.