AI chip developed for decentralized use without the cloud

(Press-News.org) A new AI chip developed at the Technical University of Munich (TUM) works without the cloud server or internet connections needed by existing chips. The AI Pro, designed by Prof Hussam Amrouch, is modelled on the human brain. Its innovative neuromorphic architecture enables it to perform calculations on the spot, ensuring full cyber security. It is also up to ten times more energy efficient. The professor of AI processor design at TUM has already had the first prototypes produced by semiconductor manufacturer Global Foundries in Dresden. Unlike conventional chips, the computing and memory units of the AI Pro are located together. This is possible because the chip applies the principle of ‘hyperdimensional computing’: This means that it recognizes similarities and patterns, but does not require millions of data records to learn.

Instead of being shown countless images of cars, as with the deep learning method used in conventional AI chips, this chip combines various pieces of information, such as the fact that a car has four wheels, usually drives on the road, and can have different shapes. Like the new chip, explains Prof. Amrouch, ‘humans also draw inferences and learn through similarities.’

An important advantage of brain-like thinking: it saves energy. For the training of a sample task, the new chip consumed 24 microjoules, while comparable chips required ten to a hundred times more energy - ‘a record value,’ notes Prof. Amrouch. ‘This mix of modern processor architecture, algorithm specialization and innovative data processing makes the AI chip something special.’

This also sets it apart from all-rounders like the chips from industry giant NVIDIA. ‘While NVIDIA has built a platform that relies on cloud data and promises to solve every problem, we have developed an AI chip that enables customized solutions. There is a huge market there.’

Neuromorphic chips: Modelled on the human brain The one square millimeter chip currently costs 30,000 euros. With around 10 million transistors it is not quite as densely packed or as powerful as NVIDIA chips with 200 billion transistors. But that is not Prof. Amrouch's primary concern. His team specializes in AI chips that perform the processing directly on site instead of having to send the data to the cloud to be processed along with millions of other data sets before being sent back again. This saves time and server computing capacity and reduces the carbon footprint of AI.

The chips are also customized for specific applications. ‘That makes them very efficient,’ says chip expert Amrouch. For example, they focus on processing heartrate and other vital data collected via smartwatch or navigation data of a drone. Because this personal and sometimes sensitive data remains on board the device, issues with stable internet connections or cybersecurity do not even arise. The chip expert is convinced: ‘The future belongs to the people who own the hardware.’

Further information:

Prof. Hussam Amrouch started his engagement at TUM two years ago. The Chair of AI Processor Design was created as part of Hightech Agenda Bayern. Further information: https://www.hightechagenda.de/ Prof. Hussam Amrouch is also active in the Munich Institute of Robotics and Machine Intelligence (MIRMI). His chip developments are relevant for health, the environment, and space. Further information on MIRMI: https://www.mirmi.tum.de/mirmi/startseite// Publications:

Sandy Wasif, Paul Genssler, and Hussam Amrouch. "Domain-Specific Hyperdimensional RISC-V Processor for Edge-AI Training." IEEE Transactions on Circuits and Systems I: Regular Papers (2025). https://ieeexplore.ieee.org/document/10931124 Soliman, Taha, Swetaki Chatterjee, Nellie Laleni, Franz Müller, Tobias Kirchner, Norbert Wehn, Thomas Kämpfe, Yogesh Singh Chauhan, and Hussam Amrouch. "First demonstration of in-memory computing crossbar using multi-level Cell FeFET." Nature Communications 14, no. 1 (2023): 6348. https://www.nature.com/articles/s41467-023-42110-y Wei-Ji Chao, Paul R. Genssler, Sandy A Wasif, Albi Mema, Hussam Amrouch, “End-to-end Hyperdimensional Computing with 24.65 µJ per Training Sample in 22 nm Technology”, under review at the European Solid-State Electronics Research Conference (ESSERC). Preprint available: https://go.tum.de/440497 Additional material for media outlets

Images for download: https://mediatum.ub.tum.de/1781785 Subjects matter expert:

Prof. Hussam Amrouch

Chair of AI Processor Design (AI-Pro)

Technical University of Munich (TUM)

München

amrouch@tum.de

TUM Corporate Communications Center contact:

Andreas.Schmitz@tum.de

presse@tum.de

END