The NeurOSmart project aims to set a new standard for intelligent hybrid computing architectures in autonomous machines and transportation systems. For this purpose, a high-performance sensor system, AI-supported preprocessing and a novel high-performance, analog-neuromorphic, ultra-low-power in-memory accelerator chip are combined. The prospect is an increase in energy efficiency of data processing by at least two orders of magnitude. This will enable the development of novel autonomous systems with previously unattainable intelligence and energy efficiency.
As mobile robot systems become more autonomous, the number of sensors increases, the effort required to link their data increases, and with it the need for computing power to realize reliable and safe real-time operation. Architecture scalability, sufficient transmission bandwidth between sensor and data processing, and minimization of power requirements are the main challenges for the development of high-performance computers to be used in mobile systems. It is predicted that in less than 10 years, the required computing capacity in the sensor periphery will have to match that of a supercomputer today. This requirement can only be met by a combination of hardware and software components specifically developed for each other.
Therefore, as part of the NeurOSmart flagship project, five Fraunhofer Institutes (ISIT, IPMS, IMS, IWU, IAIS) are researching a direct integration of the data-processing intelligence into the sensor system under the direction of the Fraunhofer Institute for Silicon Technology ISIT in Itzehoe.
This direct integration means that a significant proportion of the computing load is implemented in a particularly environmentally friendly and resource-saving manner, because the computing hardware in the sensor system can be adapted to the requirements of the sensor system directly during sensor development in the codesign. As a pioneer of integration into a competitive sensor system, NeurOSmart uses an open scanning LiDAR system developed by Fraunhofer as a basis to enable direct access to the incoming data streams. In addition, a highly scalable, analog-neuromorphic HPC chip is coupled with a sophisticated, AI-supported pre-processing pipeline to interpret the data directly at the sensor.