CaDS Seminar 2025 - Feb. 25

Dr. Rodrigo Bartolomeu (SDL Complex Particle Systems)

Performance and Portability across GPU vendors and programming models

Abstract:

The accelerated HPC landscape has changed notably in recent years, creating new challenges for scientific code developers. Nvidia was the main provider of GPGPUs for such systems in the beginning. Most Tier 1 supercomputers required the adoption of CUDA to write codes that could correctly explore accelerators' compute capabilities. However, in recent years, other vendors such as AMD and Intel have picked up on the hardware development. As of November 2024, the Top 10 in the Top 500 list have systems with hardware from the three major vendors. Maintaining entire codebases that run on several architectures is now highly desirable and laborious. The performance portability field of research addresses this by evaluating portable programming models and the expected performance one can expect by using them. In this presentation, we will share our experience and present the results of our recent study that used the same code for the N-Body problem on four different architectures from the three major vendors: Nvidia's GH200 and A100, AMD MI250, and Intel's GPU MAX 1100 in seven programming models.