A groundbreaking chip technology that combines data storage and processing has been developed, offering substantial improvements in efficiency and performance. This technology, drawing inspiration from the human brain, is anticipated to be commercially available in three to five years, but requires cross-disciplinary efforts to adhere to industry security norms.
Developed by Hussam Amrouch, a professor at the Technical University of Munich (TUM), this AI-compatible architecture outperforms existing in-memory computing methods by a factor of two. As detailed in the journal Nature, Amrouch’s innovative approach involves using a unique computational model with ferroelectric field effect transistors (FeFETs). This technology is expected to be beneficial for generative AI, deep learning, and robotics in the near future.
The principle behind this technology is straightforward: these chips not only perform computations on transistors but also store data there, leading to time and energy savings.
“This dual functionality enhances chip performance,” states Hussam Amrouch, an AI processor design expert at TUM.
These transistors, merely 28 nanometers in size, are densely packed into the new AI chips, necessitating future chips to be faster and more energy-efficient than their predecessors. Their ability to remain cool is crucial for applications like real-time drone flight calculations.
“Such tasks are highly complex and power-intensive for computers,” Amrouch adds.
Prof. Hussam Amrouch has spearheaded the development of this energy-efficient AI Chip. Credit: Andreas Heddergott / TUM
Efficiency and Power in Advanced Chips
The efficiency of these chips is quantified by the metric TOPS/W: “tera-operations per second per watt”. This represents a crucial measure for future chip performance, indicating how many trillion operations a processor can execute per second with one watt of power.
The innovative AI chip, a collaboration between Bosch, Fraunhofer IMPS, and supported by GlobalFoundries in the manufacturing process, achieves 885 TOPS/W. This efficiency is double that of similar AI chips, including Samsung’s MRAM chip, while current CMOS chips operate at 10–20 TOPS/W, as reported in a recent Nature publication.
Innovative Chip Architecture Inspired by the Human Brain
The design of this modern chip is influenced by human brain functionality. “In the brain, neurons process signals while synapses store this information,” explains Amrouch, likening it to human learning and memory.
The chip utilizes ferroelectric (FeFET) transistors, electronic switches with unique properties (polarity reversal under voltage) capable of retaining data even when disconnected from power. They also facilitate concurrent data storage and processing.
“With these, we can create highly efficient chips suitable for deep learning, generative AI, and robotics, processing data where it’s generated,” believes Amrouch.
Towards Commercially Viable Chips
The aim is to deploy these chips for complex tasks like running deep learning algorithms, object recognition, or drone data processing without delay. However, Amrouch, from TUM’s Munich Institute of Robotics and Machine Intelligence (MIRMI), estimates it will take three to five years for these in-memory chips to be ready for real-world applications.
A key challenge lies in meeting the stringent security standards of industries like automotive. Beyond reliable functioning, the chips must conform to sector-specific criteria.
“This underscores the need for interdisciplinary collaboration, combining expertise in computer science, informatics, and electrical engineering,” Amrouch emphasizes, highlighting MIRMI’s strength in this area.
Reference: “First demonstration of in-memory computing crossbar using multi-level Cell FeFET” by Taha Soliman et al., 10 October 2023, Nature Communications.
DOI: 10.1038/s41467-023-42110-y
Table of Contents
Frequently Asked Questions (FAQs) about AI Chip Technology
What is the new AI chip technology developed by Hussam Amrouch?
The new AI chip technology, developed by Hussam Amrouch from the Technical University of Munich, integrates data storage and processing in a manner similar to the human brain. This innovative approach significantly enhances efficiency and performance, making it twice as powerful as existing in-memory computing methods. It uses ferroelectric field effect transistors (FeFETs) and is expected to be beneficial for applications in generative AI, deep learning, and robotics.
How does the new AI chip improve efficiency and performance?
By integrating data storage directly into the transistors where calculations are performed, the new AI chip saves time and energy. This dual functionality enhances the performance of the chips. Each transistor measures just 28 nanometers, allowing for millions to be placed on each chip, leading to faster and more energy-efficient operations essential for applications like real-time drone flight calculations.
When can we expect the new AI chips to be market-ready?
The new AI chips, inspired by the human brain and developed for increased efficiency and power savings, are expected to be market-ready in three to five years. However, their readiness depends on meeting industry security standards and the successful collaboration across various disciplines such as computer science, informatics, and electrical engineering.
How does the new AI chip compare to current chip technology in terms of power efficiency?
The new AI chip significantly outperforms current chip technologies in power efficiency. It achieves 885 tera-operations per second per watt (TOPS/W), which is twice as efficient as comparable AI chips, including Samsung’s MRAM chip. In contrast, current CMOS chips operate in the range of 10–20 TOPS/W.
What are the potential applications of this new AI chip technology?
This AI chip technology is designed for energy-intensive applications such as deep learning, generative AI, and robotics. It is particularly suited for tasks where data need to be processed quickly and efficiently, such as object recognition, running deep learning algorithms, and processing data from drones in flight with no time lag.
More about AI Chip Technology
- Technical University of Munich
- Nature Communications
- Hussam Amrouch’s Research
- Ferroelectric Field Effect Transistors
- AI Chip Efficiency and Performance
- In-Memory Computing Technology
- Generative AI and Deep Learning Applications
- Energy Efficiency in Chip Design
- Bosch and Fraunhofer IMPS Collaboration
- GlobalFoundries Manufacturing Process
