Over the last few months, we have tested multiple different OpenMP compilere to figure out how far they have come with GPU offloading and how they perform.
We have come to the conclusion, that we cannot recommend using any version of GCC. GCC has poor support for newer GPU architectures and it delivers poor performance.
Instead our recommendation is to use an LLVM compiler from your GPU manufacturer. This means using NVIDIAs HPC SDK or AMDs ROCm stack.
20 Sep 2021