Speakers
Dr
Olivier Mattelaer
(UCLouvain/CISM)Mr
Orian Louant
(Université de Liège)
Description
This session takes the form of a serious game in which participants compete to transform a poorly performing GPU application into a highly efficient one. Through a series of challenges, they will learn how to analyze CUDA kernels, identify performance bottlenecks, port code to alternative GPU programming models, and apply optimization techniques commonly used in production HPC applications.
| Contents | Information |
|---|---|
| • Profiling and understanding an existing CUDA application • Identifying performance bottlenecks • Optimizing memory transfers and memory access patterns • Improving kernel occupancy and parallel efficiency • Reducing synchronization and communication overheads • Porting CUDA kernels to HIP • Benchmarking and validating performance improvements • Serious game challenges based on realistic scientific applications • Team competition and optimization leaderboard |
Prerequisite: • Basic knowledge of CUDA programming • Familiarity with GPU concepts and architecture • Being familiar with Linux and the command line • Basic knowledge of C or C++ Type: Hands-on serious game Target audience: Intermediate GPU programmers Must: Strongly recommended for anyone planning to optimize, maintain, or port GPU-accelerated scientific applications. |