Master's thesis: Language Models on the Resource Constrained Edge

Background. Large Language Models are powerful but resource-intensive, limiting their deployment on edge devices. This project explores a novel neural network attention mechanism—Inhibitor Transformer—designed for low-power, low-memory environments. The work is hosted by RISE Research Institutes of Sweden, a state-owned research institute that supports sustainable innovation across industry, academia and the public sector.

Description. This thesis investigates the deployment of the Inhibitor Transformer on platforms such as IoT devices, FPGAs, and RISC-V processors. The project includes hardware-specific optimizations, simulation of deployment scenarios, and benchmarking for performance and energy efficiency. A key goal is to identify the most suitable hardware platform and implement a working prototype.

Key Responsibilities

• Conduct a literature review on efficient attention mechanisms and hardware-aware model design

• Implement the Inhibitor Transformer and optimize it for selected hardware platforms

• Simulate and benchmark performance, memory usage, and energy efficiency

• Deploy a working prototype on real hardware

• Document findings in a scientific report

Qualifications

• Strong background in mathematics, machine learning, deep neural network architectures

• Experience with hardware-aware computing or embedded systems is a plus

• Proficiency in Python and frameworks like PyTorch or TensorFlow

• Familiarity with hardware platforms such as FPGAs, RISC-V, or IoT devices is advantageous

Terms
Scope: 30 hp, full-time
Location: Luleå (or remote with agreement)
Start: Flexible
Compensation: 10,000 SEK for travel, materials and the like after the project is completed and approved.

Please note: You need to have a valid student visa that allows you to study in Sweden during the thesis period.

We look forward to receiving your application!

Last day of application: July 29
Contact: Rickard Brännvall (rickard.brannvall@ri.se), Henrik Barestrand, Joakim Eriksson
Check-in questions (yes/no): 1-5 are required, 6-9 are beneficial, 10 and 11 are specifically a plus .