Next Generation Computing

Edge AI: Shrewd end devices save energy

Edge AI is designed to get artificial intelligence to where it's needed: in the end device. This calls for chips, algorithms and tools — researchers at Fraunhofer IIS are working on this future technology in the ANDANTE and TEMPO projects.

Thinking is hard work. And it's no different for com-puters. If they are to make decisions using artificial intelligence, they actually need far more power than people and animals. A specialist AI processor, then, needs 7000 times more energy than a bee's brain to recognize a flower in a picture, for example. This high energy consumption is especially problematic if the artificial intelligence is to be moved to the end devices – so to sensors which, installed in a bridge, are supposed to detect whether the tension of the structure changes, or in portable devices that can be switched on and off by voice commands. We also call this Edge AI.   

This kind of artificial intelligence especially offers a host of advantages: it will work even in places where no Internet connection is available. It is much faster than conventional AI, which sends the data to a cloud and analyzes it there. And it protects privacy, data sovereignty, security and safety, because the data is never surrendered, or disclosed only to the extent absolutely necessary. Until now though, its weakness is the huge energy consumption of the components. So how can we reduce the energy consumed by artificial intelligence to a level that opens the door to Edge AI applications? At Fraunhofer IIS, Dr. Marco Breiling and Dr. Loreto Mateu and their teams are now facing this challenge in two projects: in the TEMPO and ANDANTE projects, in which, alongside numerous other partners, the Fraunhofer Research Institution for Microsystems and Solid State Technologies EMFT and the Fraunhofer Institute for Photonic Microsystems IPMS are involved. “We are developing the energy-saving chips, hardware-aware algorithms and the associated tools that Edge AI needs. With these tools, we are already keeping an eye on the limits of the hardware for developing the algorithms and are considering them in the optimization process,” says Dr. Marco Breiling, Chief Scientist for Communication Sys-tems Research. In the TEMPO project, the research team is building on the Radar and Lidar application examples. In ANDANTE, its focus is on speech activity detection.

The chief attraction in both projects is the mix of digital and analog technologies that combines the advantages of both. “Multiplications and additions are no problem for analog computing – so for these applications we can bypass a complex digital circuit that would consume a lot of energy,” reveals Dr. Loreto Mateu, Group Manager for Smart Sensing and Electronics Research, who is responsible for the development of analog components. For the control, on the other hand, digital circuits are required, and these come from Breiling's team. Analog and digital circuits developed specifically for certain AI applications are what make Edge AI possible.