Better result using the TriMesh normal interpolation function. 16 threads, 10 bounces, -O3 optimization, render time 40 seconds.

Yeah, I have some weird normals going on, not sure what thats about.

16 threads, 10 bounces, no acceleration structures or AABB tests: 46 seconds. I suspect my normals are a little funky, but I think this was a noble attempt regardless.