During five years (2021 to 2025), the FVLLMONTI Project brought together leading European experts in nanotechnology, microelectronics, and artificial intelligence to develop a new class of ultra-efficient 3D artificial neural network accelerators.
Coordinated by Prof. Cristell Maneux, Director of the IMS Laboratory at the University of Bordeaux, the project marks a significant step forward in low-power, privacy-preserving AI computing.
Project Vision and Challenges
The project answers the growing need for real-time speech-to-speech translation; a task that today relies heavily on cloud computing. Existing systems demand constant access to large, power-hungry data centers, raising concerns about latency, data privacy, and energy consumption. These limitations make such systems unsuitable for standalone, mobile use in wearable technologies like earbuds.
To overcome these barriers, the FVLLMONTI consortium developed an Edge AI solution: a device capable of performing complete AI translation tasks locally, without any internet connection. This required a fundamental rethinking of computing architecture at both the hardware and software levels.
Key Innovation: The Neural Network Compute Cube (N²C²)
At the heart of FVLLMONTI’s technological breakthrough is the Neural Network Compute Cube (N²C²): a 3D hardware accelerator designed specifically for AI inference in ultra-constrained environments.
Architectural Features:
- 3D Compute Architecture: The N²C² employs vertical stacking to drastically reduce spatial footprint and interconnect distance, optimizing data movement and energy efficiency.
- Systolic Array Design: Processing units are arranged in a systolic array structure, enabling efficient matrix operations, ideal for neural network workloads.
- Vertical Gate-All-Around Nanowire FETs (VNWFETs): These advanced transistors enable vertical stacking and high-density integration, providing superior scalability and performance over conventional CMOS.
- Ferroelectric Memory Integration: Incorporating non-volatile ferroelectric components directly into the compute unit enhances speed and reduces energy cost by minimizing data transfer operations.
Software-Hardware Co-Design
Beyond hardware, the research team also advanced data compression techniques with ‘structured pruning’ strategies, allowing neural networks to be optimized for the N²C² hardware without compromising performance. By identifying and removing blocks of low value, the system reduces computational load, saves energy, and accelerates inference.
This co-design approach ensures that software and hardware are developed in tandem, unlocking maximum performance in real-world use cases.
Toward a New Class of AI Translation Devices
The long-term ambition of the FVLLMONTI consortium is to enable a fully autonomous, real-time speech-to-speech translation earpiece: a lightweight, energy-efficient device that functions entirely offline.
Such a device would:
- Translate speech in real time between multiple languages,
- Operate securely without transmitting data to external servers,
- Run on a small battery for extended periods,
- Fit seamlessly into a compact earbud format.
Potential applications span a wide range of sectors such as tourism, international trade, and diplomacy, where natural and efficient multilingual communication is key.
A European Collaborative Effort
FVLLMONTI is a multi-disciplinary effort uniting partners from France, Germany, Austria, and Switzerland, and reflects Europe’s commitment to digital sovereignty and privacy-respecting AI technologies. By tackling both the hardware and algorithmic challenges of Edge AI, the project positions itself at the forefront of embedded intelligent systems.
The FVLLMONTI project is funded by the European Union’s Horizon 2020 Research and Innovation Program.
Core team: Prof. Cristell Maneux, Dr. Jens Trommer, Dr. Guilhem Larrieu, Dr. Oskar Baumgartner, Prof. Ian O’Connor, Dr. Giovanni Ansaloni, Dr. Jean-Luc Rouas and Dr. Chhandak Mukherjee.
Key collaborators include teams from the IMS Laboratory, the University of Bordeaux, CNRS, École Centrale de Lyon, NaMLab, Global TCAD Solutions, LAAS-CNRS and EPFL.
Learn More
A recent outreach article published in Scientia Global outlines the project’s technical innovations and results: https://www.scientia.global/putting-ai-in-your-ears-with-3d-neural-networks/
