AMD GPUs are very powerful, but how can we use their full potential?
AMD GPUs have largely caught up with NVIDIA, and many new supercomputers today are based on AMD hardware. However, the software environment still poses a challenge, as many existing GPU codes are CUDA based and thus locked to NVIDIA GPUs. In this thesis we will port existing codes to AMD hardware. We will then benchmark the codes and optimize performance for the new hardware, ensuring that we will use it at maximum efficiency.
The goal of this master topic is to port CUDA based GPU codes to AMD using HiP and OpenCL with the objective of beating the performance of the CUDA baseline.
- Experience with C/C++
- Familiarity with GPU programming is very helpful