Leveraging ray tracing cores for particle-based simulations on GPUs

Neighbor searching is an essential and computationally heavy step in particle-based numerical methods such as discrete element method (DEM), molecular dynamics, peridynamics, and smooth particle hydrodynamics. This article presents a novel approach to accelerate particle-based simulations by leveraging ray tracing (RT) cores in addition to CUDA cores on RTX GPUs. The neighbor search problem is first numerically converted into a general ray tracing problem so that it can be possible to utilize the hardware acceleration of RT cores. A new, general-purpose RT-based neighbor search algorithm is then proposed and benchmarked with a prevailing cell-based one. As a showcase, both algorithms are implemented into a GPU-based DEM code for simulating large-scale granular problems including packing, column collapse and debris flow. The overall simulation performance is examined with varying problem sizes and GPU specs. It demonstrates that the RT-based simulations are 10%-60% faster than the cell-based ones, depending on the simulated problems and GPU specs. This study offers a new recipe for next-generation high-performance computing of large-scale engineering problems using particle-based numerical methods.

Automated summary

This article presents a novel approach to accelerate particle-based simulations by leveraging ray tracing cores in addition to CUDA cores on RTX GPUs. A new, general-purpose RT-based neighbor search algorithm is proposed and benchmarked with a prevailing cell-based one. The study demonstrates that the RT-based simulations are 10%-60% faster than the cell-based ones.