With the in-depth application of artificial intelligence (AI) in healthcare, autonomous driving, smart cities, and other fields, traditional computing architectures are facing great challenges in terms of processing speed and energy efficiency. To solve this problem, eight organizations from Europe and South Korea jointly launched an innovative project called "ViTFOX", which aims to create a new vision Transformer architecture based on ferroelectric oxides to significantly reduce computing energy consumption and latency. Unlike traditional architectures that rely on "storage-compute separation", ViTFOX achieves superior energy efficiency of more than 50 trillion operations per watt (50 TOPS/W) by embedding compute directly into storage. The project received 1.5 million euros in funding from the European Union.
In the current digital era where the amount of data is increasing dramatically, traditional computing methods are difficult to meet the needs of AI applications in terms of processing efficiency and power consumption control. Neuromorphic Computing, a system that mimics the way the human brain works, enables faster and more efficient calculations through the use of specialized hardware such as ferroelectric devices, thereby supporting real-time processing and intelligent decision-making. This new paradigm not only improves the performance of AI tasks such as image recognition and natural language processing, but also significantly reduces energy consumption, providing a sustainable solution for the development of future technologies.
The ViTFOX project is at the forefront of AI chips, advancing AI technology evolution through the development of energy-efficient brain-inspired computing platforms. Its core is a neural network architecture based on Vision Transformer, which is good at handling image recognition tasks and can extract and analyze visual information more efficiently. ViTFOX's new ViT system combined with ferroelectric oxide materials aims to achieve ultra-low-power operation in edge AI applications. Prof. Thomas Kämpfe, project leader from Fraunhofer IPMS, said: "We are exploring hardware and software co-optimization platforms, new materials and their integration technologies to improve AI performance and energy savings in computing systems. He added, "We want to make a positive contribution to the semiconductor industry and respond to the technical challenges of emerging memory technologies and the urgent need for efficient computing."
Figure: Europe and South Korea work together to create a new architecture of ultra-energy-efficient brain-like AI chips
Strengthen cooperation between South Korea and Europe to jointly promote the development of cutting-edge semiconductor technology
The ViTFOX project consists of eight leading research institutes, universities and technical laboratories from Europe and South Korea. The project aims to consolidate the leading position of the European Union and South Korea in the field of hafnium ferroelectronics compatible with silicon processes, which was first pioneered in Europe and has also been highly valued by the Korean scientific community in recent years. The project will comprehensively promote breakthroughs in the technology chain, including material and device development, heterogeneous and monolithic integration technologies, and the design and simulation of ViT system circuits.
One of ViTFOX's technical highlights is the application of recent advances in ferroelectric materials, in particular hafnium zirconium oxide (HZO). This material is not only compatible with mainstream silicon chip processes, but also shows great potential to improve memory performance and reduce energy consumption.
The three technical goals of the project focus on the design and manufacture of the core components of the vision Transformer: one is to build a demonstrator based on the concept of "in-memory computing"; The second is to develop circuit-level simulation tools; The third is to establish a software and hardware collaborative optimization platform that supports ferroelectric oxides. The platform will support two new storage technologies: a Korean-developed high-density 3D ferroelectric random-access memory (3D FeRAM) and a European-led epitaxial ferroelectric tunnel junction technology. Through the in-depth collaboration between South Korea and Europe, the project brings together multiple professional forces in materials science, semiconductor technology and AI algorithms, and is pushing this emerging field to a new stage of technology implementation.
