AI Accelerator for Energy-Efficient Edge AI

• Typ der Arbeit: Forschungsprojekt
• Status der Arbeit: offen
• Betreuer: Davis Rakhshan

TL;DR

Design and evaluate a custom AI accelerator board based on the STM32N6 for integration into an existing sensor system, enabling efficient neural-network inference directly at the edge.

Motivation

Modern sensor systems increasingly benefit from processing data directly on the device instead of transmitting raw data to an external computer or cloud service. Such Edge AI can reduce communication overhead and latency while enabling more autonomous and energy-efficient sensor systems.

The STM32N6 combines a microcontroller with dedicated hardware for neural-network acceleration, making it an interesting platform for bringing significantly more AI processing capability to embedded sensor systems.

This thesis investigates the design and integration of a custom STM32N6-based AI accelerator board for an existing sensor platform.

Research Question

How can a dedicated STM32N6-based accelerator be integrated into an embedded sensor system to enable efficient on-device AI processing?

Scope

The thesis combines PCB design, embedded-system development, and Edge AI. The exact hardware architecture and AI workloads will be refined together with the student during the project.

The work will include:

• Design of a custom PCB based on the STM32N6
• Integration of external DRAM to support larger AI workloads
• Design of power supply, communication, programming, and debugging interfaces
• PCB layout, manufacturing, assembly, and hardware bring-up
• Firmware development and integration with the existing sensor system
• Deployment of representative neural-network models to the AI accelerator
• Evaluation of inference performance, memory usage, and energy consumption
• Analysis of the benefits and limitations of dedicated Edge AI processing within the sensor system

The project should fulfil agreed milestones covering hardware design, board bring-up, system integration, AI deployment, and final performance evaluation.

Expected Outcome

At the end of the thesis, a functional AI accelerator board should be integrated with the existing sensor system and capable of performing neural-network inference on-device.

The work should provide an experimental evaluation of the resulting system in terms of inference performance, memory requirements, and energy consumption, as well as identify the practical benefits and limitations of the proposed architecture for Edge AI applications.

Required Skills

Required

• Basic understanding of digital electronics and embedded systems
• Experience with C/C++ or similar low-level programming
• Basic understanding of microcontrollers and digital interfaces
• Interest in PCB design and hardware development
• Basic understanding of machine learning and neural networks
• Ability to work independently and document results clearly

Desirable

• Experience with STM32 microcontrollers
• Experience with KiCad, Altium Designer, or similar PCB design tools
• Experience with PCB assembly and hardware debugging
• Familiarity with external memory interfaces such as DRAM
• Experience with PyTorch, TensorFlow, ONNX, or similar frameworks
• Familiarity with model quantisation and embedded AI deployment
• Experience with oscilloscopes, logic analysers, or power measurement equipment
• Experience with Git and LaTeX

Prior experience with all of the desirable skills is not required. Relevant PCB design, embedded-system, and Edge AI knowledge can be acquired during the project.

Deliverables

• Written thesis
• Custom STM32N6-based AI accelerator PCB
• Reproducible hardware design and firmware
• Demonstration and experimental evaluation of on-device AI inference
• Final presentation

Application

Please include:

• Curriculum vitae
• Current transcript of records
• Short description of your motivation and relevant experience

Please keep your application concise. A CV, transcript of records, and a short description of your motivation and relevant experience are sufficient. Lengthy motivation letters or extensive application documents are neither required nor encouraged.