From Threads to the AI Age: Rethinking GPU Programming

Abstract: Artificial intelligence is reshaping many industries, driven by GPUs in data centers that provide crucial computing power. The immense scale of today’s workloads makes it vital to create highly efficient kernels—a complex challenge. This presentation will track how GPGPU programming has evolved: from early thread-based approaches and tile-focused specialized languages, to the rise of code generated by AI. We'll discuss fresh innovations that fuel ongoing progress, including smarter libraries and new programming models for GPUs, which prioritize data locality and improve dataflow. As AI-powered systems start automating and streamlining code optimization, we are approaching an era where GPU acceleration becomes more efficient and far easier to access. How should programming evolve for a world where machines do the heavy lifting? 

Bio: Samuel Bayliss is a Fellow within the Research and Advanced Development (RAD) group at AMD, where he leads initiatives focused on developing advanced programming tools for AMD’s GPUs and machine learning accelerators.

He brings extensive expertise in compilers and computer hardware, having contributed to CPU development at ARM and worked on ML accelerators and compiler technologies at Xilinx and AMD. His team is dedicated to advancing open-source compiler tooling through innovative MLIR abstractions and integrating hardware architecture design with software tools to optimize the efficiency of machine learning systems. 

Dr. Bayliss earned his M.Eng. and Ph.D. degrees from Imperial College London, specializing in high-level synthesis tooling and memory optimization for accelerator architectures.