Hardware Ray Tracing BVH (Bounding Volume Hierarchy) Traversal Optimization: Advanced Implementation and Real-World Architecture

I’ve been deep in the weeds optimizing my BVH traversal pipeline for real-time ray tracing and hitting some real bottlenecks. My ray-box intersection tests are running cheap, but ray-triangle tests are eating my budget—specifically the Möller-Trumbore rejection path. For anyone who’s actually implemented this on the GPU or even done CPU-level work: how are you handling the refit versus rebuild decision when objects move? I keep seeing the claim that refitting is cheaper, but in practice I’m seeing refits blow past rebuild costs once you hit a certain triangle velocity threshold. What’s your actual tipping point, and how are you measuring it?

On the architecture side, I’m curious whether people are keeping their BVH traversal in the shader or pushing it to a dedicated compute pass. I’ve read that hierarchical traversal with early ray-BOX rejection on the parent nodes saves a ton of triangle tests, but I’m not seeing the expected speedup in my benchmarks—my refit overhead is dominating when dynamic geometry is involved. Are you using surface node refitting, or do you just rebuild per frame for dynamic scenes? And critically, what’s your measured cost delta between the two at, say, 60k triangles with partial updates?