Hardware Agnostic Artificial Intelligence for Embedded

01/04/2021 to 30/05/2023

In embedded computing, existing neural networks should be able to be deployed on different hardware and produce better, the same or similar comparable results regarding scaled performance, latency, power and results without needing a software developer to re-write their applications. The University of Tübingen is developing novel approaches to evaluate and optimize different neural processing units (NPUs) running with ARM-based hardware; and to demonstrate the strength of the open-source machine learning compiler framework TVM or SiMa.AI SDK and show to what extent deep neural networks can be developed NPU-independently. Furthermore, the University of Tübingen develops and applies representative machine learning workloads for different application domains and ports these workloads onto different NPUs.