PhD projects at SDU Microelectronics
For more information
Ongoing Projects
Alexander Borch Kristensen
Project: NeuroEar - In-Ear Multi-Modal Monitoring of Brain Disorders
Supervisor: Associate Professor Milad Zamani
With its scalability, high temporal resolution, and portability, electroencephalography (EEG) is one of the most popular modalities for recording instantaneous physiological activity in the brain. Transcranial bioimpedance (BioZ) as a modality offers physical characterization of the underlying tissue, reflecting instantaneous cerebral components information. The NeuroEar project aims to combine both modalities into one discreet in-ear recording device, with low power consumption, thereby offering long-term combined physiological and physical characterization for research in various brain disorders.
Ming Zhang
Project: Energy-Efficient and Low-Latency SSM-Transformer Accelerator for LLMs
Supervisor: Professor Farshad Moradi
Although transformer-based large language models (LLMs) achieve impressive results, their quadratic computational complexity leads to significant latency and high energy consumption, making deployment on edge devices challenging. State Space Models (SSMs) and hybrid SSM-Transformer architectures offer linear time complexity and reduced memory requirements, but conventional hardware accelerators do not efficiently support SSM-specific execution patterns such as selective scanning and recurrent updates. This results in significant inefficiencies in dataflow and scheduling. To address this challenge, the project proposes an algorithm-hardware co-designed accelerator targeting hybrid SSM-Transformer LLMs. The accelerator employs a unified dataflow, reconfigurable computing units, and a hierarchical memory system to enable ultra-low-latency and energy-efficient implementation of LLMs on resource-constrained platforms.
Yaogang Wang
Project: Low-Power Architectures for Multi-Channel Intelligent Neural Interfaces
Supervisor: Professor Farshad Moradi
The main objective of the project is to develop an energy-efficient multi-channel neural interface for recording data from multiple channels. The system should be configurable for different applications while also being able to adapt quickly and efficiently to different scenarios in a flexible and dynamic manner. The aim is to minimize the system's energy consumption while optimizing other relevant performance parameters. The project therefore focuses on developing intelligent and energy-efficient architectures for multi-channel neural interfaces.