Cosmological Simulations on GPUs in the Era of High-Precision Cosmology

In recent years, the role of Graphics Processing Units (GPUs) has become increasingly relevant in the world of high-performance scientific computing. In the context of large cosmological simulations, where several million hours of computational time are needed, the use of GPUs enables running bigger simulations at higher resolution. In this master’s thesis, we study the current implementation of the gravity solver in the cosmological simulation code, with the goal of assessing its numerical accuracy across different hardware architectures and simulation setups. Using dark-matter-only simulations, using initial conditions from the Magneticum simulations, we analyse a set of key cosmological observables, including the matter power spectrum, the halo mass function, and the internal properties of virilised haloes. Particular attention is devoted to a detailed investigation of the treePM scheme employed by OpenGadget3. By exploring the effects of different configuration flags and numerical parameters, we quantify their impact on cosmological observables at the percent-level.