October 12, 2023 — Ben van Werkhoven gave a talk titled “Automatically optimizing GPU applications for performance and energy efficiency” at the Bits&Chips Event in Eindhoven, the Netherlands. The Bits&Chips Event brings together technical management, engineers and researchers in complex software and system development.
Graphics Processing Units (GPUs) are fueling the current AI revolution and have dramatically reshaped the computing landscape over the past two decades. Serving as the primary driving force in training large language models, such as ChatGPT, as well as many large-scale scientific applications, GPUs have become the go-to platform for high-throughput and energy-constrained computing. However, unlocking the full computational power of the GPU is a significant challenge. Reducing the carbon footprint and optimizing the performance of GPU applications require exploration of vast and discontinuous program design spaces, which is an infeasible task for programmers to perform manually. Moreover, this search process needs to be repeated for different hardware and for various input problems, leading to productivity and maintainability challenges. This talk provides an overview of our research into Kernel Tuner, a software development tool that enables GPU programmers to create GPU applications with optimal compute and energy performance.
