In the present manuscript ‘A GPU-parallelization of the neXtSIM-DG dynamical core (v0.3.1)’ the authors test and evaluate different GPU programming frameworks based on their sea ice model dynamical core neXtSIM-DG.

Many modeling groups in the weather and climate community and beyond are facing similar problems as the neXtSIM-DG developers. Developing portable code that achieves good performance on various hardware architectures without limiting the productivity of the (scientific) developers too much is a major challenge. Therefore, the thorough analysis of the different available GPU programming frameworks presented here is of great value to the community. The study is well written and I would recommend publication in GMD after a few issues have been addressed as listed below.

1. In line 370 in section 4.1 it is stated ‘These results indicate that the AMD ecosystem is still less mature’. However, to validate this statement and to have a complete picture also for AMD GPUs it would have been nice to also have a HIP implementation as a baseline to compare the other implementations against similar to the CUDA implementation for NVIDIA GPUs.
2. Table 1: What hardware was used for these measurements and how many OpenMP threads were used?
3. Figures 2 and 10 and lines 354 and 470: Again, how many OpenMP threads were used for the OpenMP reference simulation? And what backend was used for Kokkos on CPUs? OpenMP as well? And if yes, with the same number of threads as the reference OpenMP simulation?
4. Line 206: LLVM/Clang provides a set of debugging flags (e.g. https://openmp.llvm.org/design/Runtimes.html#libomptarget-info) which can provide precise information about each block of memory and potential problems. Also, for the types that are not trivially copyable, OpenMP 5.0 offers the option of using declare mapper to define this. Wouldn’t that have been an option here?

Technical corrections:

• Table 2: ‘AdaptiveCPP’ is used here to indicate the SYCL implementation but the name is too generic. AdaptiveCPP is also the name of the compiler and it can also compile native OpenMP or other parallel APIs. I would suggest replacing ‘AdaptiveCPP’ with ‘SYCL-AdaptiveCPP’.
• Figure 2: Why is in the legend of the right panel TorchInductor marked with an ‘*’?
• Line 16: impact on long-term processes
• Line 42: is -> it
• Line 88: often often -> Remove one