The thesis explores the potential of efficient resource management and real-time scheduling strategies for deploying advanced deep neural networks (DNNs), such as transformers, on low-power embedded systems. Possible areas of focus include creating real-time scheduling algorithms for timely and efficient AI inferences, optimizing the use of low-power AI accelerators, improving SRAM and memory bandwidth utilization, and implementing dynamic power and thermal management techniques. Practical applications like virtual reality and real-time text generation will be considered to validate the effectiveness of these methods, aiming to enhance the deployment of efficient and reliable AI solutions in resource-constrained environments.

Feel free to contact Binqi Sun (binqi.sun@tum.de) for latest available topics. You are also welcome to bring new ideas/challenges!

 

Hands-on experience with DNN tranining and inference. 

Basic knowledge of combinatorial optimization and scheduling.