AI accelerators have proliferated in data centers in recent years and are now almost ubiquitous. In addition, their computational power and, most importantly, their energy efficiency are up to orders of magnitude higher than that of traditional computing. Over the last years, various methods and optimizations have been tested to use these hybrid systems for simulations in the context of astroparticle physics utilizing CORSIKA. The main focus of this talk is the propagation of optical, i.e. fluorescence and Cherenkov, photons through low density inhomogeneous media in the context of the next generation CORSIKA8 simulation framework. Different techniques used and approximations, e.g. the atmospheric model, tested during the development will be presented. The trade-off between performance and precision allows the experiment to achieve its physical precision limited to the real resolution of the experiment and not invest power and time in vanishing precision gains. The additional comparison of classical CPU-based simulations with the new methods validates these methods and allows evaluation against a known baseline.