The KI‑FLEX project set out to create a low‑power, high‑performance ASIC that could run the latest deep‑learning algorithms required for fully automated driving while consuming less than 10 W. The design goal was to replace bulky, energy‑hungry GPUs and large FPGAs with a flexible, multi‑core accelerator that could be integrated into vehicles at the time of production and still execute future, unknown neural‑network concepts. To achieve this, the consortium developed a complete hardware and software stack. The ASIC was designed to support the TensorFlow API, allowing the same neural‑network models that were trained on conventional platforms to run directly on the chip. Training and inference pipelines were built for camera‑based object detection, semantic segmentation of images and point clouds, and for fusion of LIDAR and camera data. A semantic SLAM module was added that uses the segmentation labels to improve localization against high‑definition maps. The system also incorporates a modular communication framework that enables future automotive architectures to plug in new sensors or algorithms without redesigning the core.

Performance results reported in the final report highlight the ASIC’s low power envelope and its ability to execute complex DNNs. The chip’s power consumption was measured below the 10 W target, a significant improvement over comparable GPU or FPGA solutions. While the report does not provide explicit accuracy figures, it states that the neural‑network models for static and dynamic object detection, as well as for semantic segmentation, achieved “equal or higher accuracy with lower energy consumption” compared to baseline implementations. The ASIC was validated on a reconfigurable FPGA prototype before final silicon fabrication, ensuring that the design met the required throughput and latency for real‑time perception. Additional validation was performed on solid‑state LIDAR sensors, demonstrating the system’s capability to handle emerging sensor technologies.

The project was organized into three work packages: AP1 defined requirements and concepts, AP2 focused on component development, and AP3 handled system integration and testing. The consortium comprised a mix of research institutes, universities, and industry partners. Fraunhofer institutes (IPMS, EMFT, IIS), the Technical University of Dresden, and Friedrich-Alexander University Erlangen contributed expertise in digital circuit design, signal processing, and automotive software. Industry partners such as Valeo, Heimann Sensor, and eesy‑innovation supplied sensor hardware, automotive integration experience, and market insight. The collaboration also linked to the ANDANTE project, which pursued dedicated neural‑network hardware, allowing cross‑fertilization of design ideas and validation methods.

Funding for KI‑FLEX came from a German federal grant, supporting the consortium’s research and development activities over the project’s duration. The final report emphasizes the commercial relevance of the results: the developed ASIC, along with its software stack, can be incorporated into future automotive production lines, providing robust perception, localization, and mapping capabilities while keeping power budgets low. Planned publications include a contribution to the International Journal of Transportation and a short technical report, ensuring that the scientific community can assess the methods and results. Overall, KI‑FLEX delivers a scalable, energy‑efficient deep‑learning accelerator that aligns with the automotive industry’s move toward dedicated AI hardware for autonomous driving.