Energy-efficient AI with ferroelectric technology

Conventional computer architectures are limited in terms of processing speed and energy efficiency when dealing with the huge amounts of data generated in today's digital landscape. Neuromorphic systems, systems that mimic the way the human brain works, use specialized hardware, such as ferroelectric devices, to perform computations more efficiently, enabling real-time processing and decision making. This approach not only increases the performance of AI applications such as image recognition and natural language processing, but also reduces energy consumption, making it a sustainable solution for future technologies.

The 'Vitfox' project brings together eight partners from Europe and Korea to develop a vision transformer architecture based on ferroelectric oxide that enables a significant reduction in energy consumption and latency. In contrast to conventional architectures, which often rely on separate memory and processing units, Vitfox aims to integrate data processing directly into the memory and thus achieve an improved energy efficiency of 50 TOPS/W. The European Union is funding the project with 1.5 million euros.

At the heart of the project is the Vision Transformer (ViT) architecture, which is designed to perform complex AI calculations while consuming less energy. Vision Transformers are a type of neural network architecture that processes visual data more effectively than conventional methods. The project aims to develop a ViT that utilizes ferroelectric oxide materials to achieve an energy efficiency of over 50 TOPS/W. This threshold is critical for AI-powered edge applications. "We aim to push the boundaries of current technology by developing hardware-software co-optimization platforms, novel materials and integration methods. These can not only improve AI performance, but also ensure sustainable energy consumption," says Prof. Dr. Thomas Kämpfe, project manager at Fraunhofer IPMS, one of the partners in the consortium. "We want to make an important contribution to the semiconductor industry by addressing both the technical challenges of new memory technologies and society's need for efficient computing solutions," he adds.

Cooperation between Europe and Korea

In total, the Vitfox consortium consists of eight partners from leading research institutions, universities and technology development laboratories from Europe and Korea. The project aims to strengthen the position of the EU and Korea in the field of silicon-compatible ferroelectric electronics based on hafnium. The field has already seen much pioneering work within Europe and has now attracted great interest from Korean researchers. The project will advance the technology beyond the current state, along the entire value chain from materials and devices, to heterogeneous and monolithic integration and the design and simulation of ViT circuits and systems. The project is of particular importance as it benefits from recent advances in ferroelectric materials, especially hafnium zirconium oxide (HZO). As it has proven to be compatible with conventional silicon devices, it represents a promising approach for improving memory devices and reducing power consumption.

Three of the project goals are focused on the development and production of the main components of the ViT. A compute-in-memory demonstrator, a circuit-level simulator and a platform for joint optimization of hardware and software with ferroelectric oxides are planned. The platform will support two types of new memories: the high-density 3D FeRAM developed in Korea and the epitaxial ferroelectric tunnel junctions developed in Europe. The close collaboration will enable the partners to combine their joint expertise in materials science, semiconductor technology and artificial intelligence to push the boundaries of this emerging field.