Accelerators are essential for achieving optimal performance and energy efficiency in computing. However, market segmentation often leads to language lock-ins, limiting flexibility across accelerators and increasing development costs. To address this, alternatives like SYCL or OpenMP have emerged, enabling code portability across  diverse hardware platforms.
The study reveals that while both OpenMP and SYCL demonstrate comparable performance to native languages on NVIDIA and Intel GPUs, SYCL significantly outperforms OpenMP on AMD platforms. These findings underscore the potential of multi-device and open languages in enhancing performance and reducing development overhead in parallel CFD simulations.