Increasing simulation throughput is a major challenge for LHC experiments as they undergo significant detector upgrades for the high-luminosity phase. GPU-enabled particle transport simulation is a key R&D direction to address this, leveraging the growing availability of GPUs in computing centers. In its first phase, the AdePT project demonstrated that particle transport codes can be adapted for GPUs, integrated into a standard Geant4 workflow, and deliver significant speed-ups for standalone Geant4 setups of varying complexity. The second phase focuses on enabling seamless and efficient GPU usage within experiment frameworks via a Geant4 plugin. To achieve this, GPU transport kernels have been restructured into a header library hidden from the users, exposing only a configurable integration library easy to interface from diverse Geant4 applications. Several performance limitations identified in the first phase have been partially addressed. CPU-GPU scheduling has been improved to process multiple events on the GPU while allowing the CPU to perform asynchronous tasks. In addition, we continued the development of a new GPU-friendly surface-based geometry model, which mitigates some of the geometry-related bottlenecks. The initial integration of AdePT with two experiment frameworks has revealed challenges that will be addressed moving forward. Here, we present the latest results and insights, focusing on the hybrid Geant4-AdePT use case.
Apostolakis et al. (Tue,) studied this question.