Kokkos 3 introduces new abstractions like hierarchical parallelism, advanced atomics, and graph constructs to improve performance portability for exascale systems. Benchmarks demonstrate the performance benefits of these features on both CPUs and GPUs, showcasing Kokkos's ability to adapt to diverse architectures.

The authors are affiliated with multiple national laboratories (Sandia, Oak Ridge, Lawrence Berkeley, Argonne, and Swiss National Supercomputing Centre) which are involved in the development and funding of Kokkos. This could potentially introduce bias in the presentation of results and comparisons.

This paper describes valuable extensions to the Kokkos programming model that address challenges in exascale computing. The introduction of hierarchical parallelism, advanced reductions, arbitrary-sized atomics, and graph constructs enhances the model's expressiveness and performance potential. While reproducible benchmarks are provided, the limited hardware specifics and lack of thorough comparisons with other performance portability models slightly lower the rating. The potential conflict of interest arising from authors' affiliations with Kokkos-involved institutions is acknowledged.