Gábor is a researcher, developer and university lecturer specialising in computer graphics and geometric algorithms. Earlier in his career, he spent four years working in AAA game development at Visual Concepts on the NBA 2K series. He later spent five years at Shapr3D, where he worked as a Rendering Engineer and Engineering Manager, leading rendering teams and working on new platforms such as Apple Vision Pro. He currently teaches and conducts research at the Faculty of Informatics at ELTE, focusing primarily on real-time graphics, geometric modelling, distance fields and robust geometric algorithms running on GPUs. His work focuses on turning mathematically fascinating ideas into graphics solutions that are genuinely fast, practical and useful in real-world applications. Talk: Shader Optimisation Beyond the Compiler: AMD and NVIDIA ISA Assembly in Practice This talk explores how modern GPUs actually execute carefully written shader code. We will look at AMD RDNA architectures, which also form the basis of the current console generation (PS5 and Xbox Series X), as well as NVIDIA’s Ampere architecture, one of the most widely used GPU architectures in PC gaming. We will examine the key similarities and differences in their hardware design, then use low-level, so-called ISA assembly to trace how operations in high-level shader languages are translated into the elementary instructions actually executed by the GPU. The aim is not to teach assembly programming, but to develop a more accurate mental model of what happens between shader source code and the programme running on the hardware: how vectorisation, instruction selection, register usage, memory access, parallel execution and control-flow divergence manifest at the hardware level. This perspective can help identify performance issues that may not be apparent from the source code alone, and explain why the same shader can behave differently on AMD and NVIDIA hardware. Finally, we will look at how this knowledge can be applied to AI-assisted programming: how to verify the real cost of generated code, how to give code-generation tools more precise optimisation constraints, and how to distinguish changes that merely look convincing from those that deliver measurable performance improvements.