This works for triangles in pre-defined scenes where we can compute the BVH in advance, but how does it work in practice for dynamically generated triangles? Will it be less efficient if we still generate the BVH in real-time before calculating the ray-triangle intersection?

Maybe this paper will interest you! https://research.nvidia.com/sites/default/files/pubs/2010-06_HLBVH-Hierarchical-LBVH/HLBVH-final.pdf