Celeritas: Using GPUs to accelerate HEP detector simulation

Celeritas is a new Monte Carlo (MC) detector simulation code that allows experiment workflows to achieve higher performance (events/s) and efficiency (events/W) than current High Energy Physics (HEP) MC tools by running on GPUs, which are increasingly common in high performance computing (HPC). The latest release of Celeritas implements standard electromagnetic (EM) physics for photons, electrons, and positrons on GPU and CPU. On the Perlmutter HPC system, an EM-only idealized tracker test problem shows a single Nvidia A100 GPU to have the same throughput as 166 AMD EPYC CPU cores and is estimated to be about 3 times more power efficient than CPU. To maximize adoption by experimental collaborations, Celeritas integrates easily into existing Geant4 applications, "offloading" EM particles to accelerate the simulation. We will review these capabilities and EM performance results, and introduce new optical physics simulation capabilities being added to Celeritas to support dark matter and neutrino experiments.