The Cosmic Dawn marks the epoch in the Universe's history when the first generations of stars and galaxies emerged, fundamentally altering the state of the surrounding intergalactic medium. A primary challenge in modern cosmology is understanding the formation of these earliest objects, a process governed by a delicate and complex feedback loop: molecular hydrogen acts as the coolant necessary for gas to collapse and form stars, yet this same gas is sensitive to ultraviolet radiation—which may be emitted by the stars themselves and other sources—which dissociates the hydrogen and can suppress subsequent star formation. Understanding this interplay is essential for interpreting observations such as those by the James Webb Space Telescope (JWST). To study the formation of the first galaxies, we combine cosmological structure formation simulations with radiative transfer calculations. We use the GPU-accelerated N-body code PKDGRAV3 to model the growth of dark matter structure and identify potential source populations, while the radiative transfer code pyC2Ray follows the propagation of radiation through the intergalactic medium. Since the physical properties of the first galaxies remain uncertain, multiple source and feedback models must be explored. The primary purpose of this allocation is to establish the computational framework required for future large-scale simulations of the Cosmic Dawn. We will benchmark PKDGRAV3 to determine the computational resources required to resolve the minihalo population responsible for the earliest episodes of star formation and characterize its performance on the new Arrhenius system based on NVIDIA Grace Hopper Superchips. In parallel with the benchmarking of pyC2Ray on Arrhenius, we will evaluate the performance and portability of a newly developed HIP implementation on AMD GPUs through Dardel-GPU. The project brings together the core pyC2Ray development team, including Prof. Garrelt Mellema (Stockholm University), Dr. Michele Bianco (Stockholm University), and the Principal Investigator (PI). We are also joined by PhD student Viktor Köhlin Lövfors (supervisors: Prof. Mellema and the PI). These studies will determine the computational requirements for future large-scale production simulations and establish performance expectations across both NVIDIA and AMD GPU architectures.