ZuSE-KI-mobil Software Development Kit for AI Acceleration
Conference: MikroSystemTechnik KONGRESS 2025 - Mikroelektronik/Mikrosystemtechnik und ihre Anwendungen – Nachhaltigkeit und Technologiesouveränität
10/27/2025 - 10/29/2025 at Duisburg, Germany
doi:10.30420/456614025
Proceedings: MikroSystemTechnik Kongress 2025
Pages: 5Language: englishTyp: PDF
Authors:
Friedrich, Martin; Lueders, Matthias; Renke, Oliver; Weddige, Sousa; Riggers, Christoph; Blume, Holger; Friedrich, Simon; Mojumder, Shaown; Matus, Emil; Fettweis, Gerhard; Ahmadifarsani, Samira; Kontopoulos, Leonidas; Schlichtmann, Ulf; Hoefer, Julian; Schmid, Patrick; Toto-Kiesa, Hella; Becker, Juergen; Kock, Markus; Schewior, Gregor; Blume, Steffen; Grantz, Darius; Benndorf, Jens; Fasfous, Nael; Mori, Pierpaolo; Voegel, Hans-Joerg; Teepe, Gerd; Bierzynski, Kay
Abstract:
AI deployment in autonomous systems demands specialized solutions meeting strict performance, energy efficiency, and cost constraints. This paper presents the ZuSE-KI-mobil Software Development Kit (SDK), a comprehensive framework for the ZuSE KI-mobil heterogeneous SoC platform. The SoC integrates dual-core Arm processors with two AI accelerators: the SPA-ML (6 TOPS) and Cadence NNA110 (4 TOPS), delivering 10 TOPS at int8 precision. The SDK enables efficient utilization and mapping through optimized runtime libraries, specialized drivers, and neural network compilers, including a custom TVM-based toolchain for SPA-ML. Four case studies demonstrate the SDK’s capabilities: camera-based 3D car detection, radar object detection using PointPillars, disengagement detection, and gesture recognition. These applications showcase flexible workload mapping across heterogeneous compute units, achieving a high performance while maintaining energy efficiency. The SDK provides a complete Yocto Linux-integrated development environment, enabling rapid prototyping and deployment of AI applications in resource-constrained autonomous systems.