Abstract

Our recent work (2509.13307) demonstrated the performance of GPU -accelerated nested sampling for efficient high-dimensional Bayesian inference in cosmology. Using JAX -based emulators and likelihoods we can leverage the parallel computing of GPUs to achieve orders of magnitude speed-ups against CPU -based analyses, and bring robust evidence calculations up to GPU -speed. This puts nested sampling back on equal footing with Markov Chain Monte Carlo (MCMC) methods, which use auto-diff gradients to achieve their speed-ups. In particular a Euclid-like mock Cosmic Shear likelihood has been considered, an analysis which previously took 8 months on a CPU instance, and we bring the time to constrain both ΛCDM and w0wa down to only 2 days on a single GPU . This talk explores a few options for pushing our methodology even further, in preparation for joint analyses of next generation of cosmlogical surveys.