Speaker
Description
The integration of AI into cosmological research is poised to significantly impact major experiments such as Simons Observatory, LiteBIRD, and CMB-S4. These projects aim to achieve unprecedented precision in mapping the cosmic microwave background (CMB), necessitating high-resolution simulations to interpret the data accurately. AI techniques, particularly those enhancing low-resolution simulations to high-resolution outputs, are instrumental in meeting these requirements efficiently.
A Super Resolution (SR) model based on the Diffusion approach has been used for PySM simulations on the SDSC Voyager system. PySM is an astrophysics package used by cosmologists to simulate full-sky maps of Galactic microwave emissions and polarization, primarily for CMB experiments. This SR model is intended to be used to enhance the resolution of images of galactic dust emissions generated by PySM. Voyager is a category II NSF system for which main components, the Intel Gaudi accelerators, were specifically designed to accelerate and scale Deep learning applications.
In this work, we present how an existing open source SR model was customized to take as input PySM images. We show the process to port it to Voyager and its Intel Gaudi architecture and software. Then we describe the path to deploy it on Voyager using Kubernetes objects. To accelerate model training, we parallelized it using Pytorch DDP to run on multiple Intel Gaudi cards. Results from a scaling test will be exhibited to discuss multi-node training on Voyager.
| Do you plan to submit a 4-page extended abstract on OpenReview (only for Presentations/Posters)? | No |
|---|