CS MSc Thesis Presentation 29 April 2026

Wednesday, 29 April there will be a master thesis presentation in Computer Science at Lund University, Faculty of Engineering.

The presentation will take place in E:2405 (Glasburen).

Note to potential opponents: Register as an opponent to the presentation of your choice by sending an email to the examiner for that presentation (firstname [dot] lastname [at] cs [dot] lth [dot] se). Do not forget to specify the presentation you register for! Note that the number of opponents may be limited (often to two), so you might be forced to choose another presentation if you register too late. Registrations are individual, just as the oppositions are! More instructions for opponents are found here on the LTH thesis project page.

09:15-10:00 in E:2405 (Glasburen) N.B. No more opponents for this presentation

• Presenter: Henning Tollin, David Stålmarck
• Title: Quantization of CNNs & LLMs for integer-only hardware
• Examiner: Flavius Gruian
• Supervisor: Jonas Skeppstedt (LTH), Davide Grohmann (Arm)

Deploying modern machine learning models on edge devices is challenging due to limited compute, memory, and energy resources. Quantization helps reduce these requirements by representing model values with lower numerical precision, but some deployment pipelines still rely on floating-point operations during inference.
This thesis explores how post-training quantized machine learning models can be converted to run using only integer arithmetic. A compiler-based workflow is developed to translate quantization parameters into an integer multiplier-and-shift representation, enabling floating-point scaling operations to be replaced with fixed-point integer computations during inference.
The work also investigates efficient integer implementations of the softmax operator, which is essential in transformer models. In particular, a new approach called DIGmax is proposed to approximate softmax using integer lookup tables.
The results demonstrate that fully integer inference can closely match floating-point behavior while enabling more efficient deployment of machine learning models on resource-constrained hardware.