AI Chip Designed for Decentralized Operation Without Relying on the Cloud

Researchers at TUM have developed the AI Pro, a neuromorphic AI chip that processes data locally, without cloud or internet reliance. Designed by Prof. Hussam Amrouch, it boosts cybersecurity and offers up to 10 times greater energy efficiency.

The professor of AI processor design at TUM has partnered with semiconductor manufacturer Global Foundries in Dresden to produce the first AI Pro prototypes. Unlike traditional chips, the AI Pro integrates computing and memory units in one place. It uses “hyperdimensional computing” to identify patterns without requiring large training data.

Pattern Recognition Without Big Data

Unlike conventional AI chips that rely on deep learning and require exposure to thousands of car images, this chip learns by combining key features—such as a car having four wheels, typically driving on roads, and coming in various shapes. As Prof. Amrouch explains, the chip mirrors human learning by drawing inferences and recognizing similarities.

One key benefit of brain-inspired processing is its energy efficiency. In a sample training task, the new chip used just 24 microjoules—10 to 100 times less energy than similar chips. “That’s a record,” says Prof. Amrouch. He highlights that the chip’s uniqueness comes from its combination of advanced processor design, tailored algorithms, and novel data handling.

This distinguishes it from general-purpose chips made by industry leaders like NVIDIA. “NVIDIA offers a cloud-based platform aimed at addressing a wide range of problems,” says Prof. Amrouch. “In contrast, we’ve created an AI chip tailored for specialized, on-site solutions—a market with significant potential.”

Research Under Review for ESSERC 2025 Presentation

“The reviewers are currently assessing the research for presentation at the European Solid-State Electronics Research Conference (ESSERC 2025).“

The current cost of the one-square-millimeter chip is 30,000 euros. Although it contains around 10 million transistors—far fewer than the 200 billion in NVIDIA chips—this isn’t Prof. Amrouch’s main focus. His team specializes in AI chips that process data locally, eliminating the need to send it to the cloud for processing alongside millions of other data sets. This approach saves time, reduces server load, and lowers the carbon footprint of AI.

The chips are tailored for particular applications, making them highly efficient, according to chip expert Amrouch. For instance, they specialize in processing data such as heart rate or other vital signs from smartwatches, or navigation data from drones. Since this personal and often sensitive data stays on the device, problems with internet connectivity or cybersecurity are avoided. “The future belongs to those who own the hardware,” Amrouch asserts.

Read the original article on: Techxplore

Read more: The Fastest AI Chip will Greatly Accelerate AI Progress