The project aims to develop an AI accelerator with a flexible, expandable, and scalable system-on-chip (SoC) architecture. For low energy consumption, the accelerator is optimized for AI algorithms in autonomous driving and manufactured in the energy-efficient 22 nm FDX semiconductor technology. In addition, Zuse-KI-mobil establishes an ecosystem that combines a development system and a German partner network with expertise in AI hardware design. Finally, the flexibility and scalability of the processor platform will be verified using demonstrators.
In Zuse-KI-mobil, a flexible and scalable processor platform with energy-efficient semiconductor technology for resource-intensive artificial intelligence (AI) applications is developed across all performance classes. It combines high computational power at low energy consumption with flexible system-on-chip architecture, enabling high-performance, scalable and application-oriented development of future AI hardware. What will set the new platform apart from state of the art is its ratio of computing power to electrical power dissipation, which with scaling effects is expected to be 300 to 400 TOPs with power consumption well below 100 W. Initially, a 22 nm semiconductor technology (22FDX from GlobalFoundries) will be used, which enables high energy efficiency at moderate cost. Smaller structure sizes to 8 nm and other semiconductor technologies/processes are also being considered to scale the performance data. The development focus is on a novel AI hardware accelerator that enables efficient, optimized, application-specific total solutions through adapted hardware/software co-design.
Furthermore, the project establishes an ecosystem for the broad deployment of the SoC. For multidimensional optimization of the architecture at all levels of energy efficiency and functional safety, novel neural networks are designed for sensor data fusion for object detection. Data from multiple cameras, lidar, and radar sensors are advantageously fused depending on the situation using advanced machine learning methods. In this way, the robustness under different weather conditions and partial failures increases, and the general detection performance is further enhanced. In addition to the use for automated driving, use cases, e.g., drones and industry, are also considered. The project results will be presented in the context of demonstrators.