What’s the issue with an academic license?

Tasty too

Couldn’t get it faster with quantization or other tricks. I am also sorting the 8 nodes by distance to AABB. Both ray and cone performance has improved dramatically (traversal is virtually negligible compared with intersecting meshes), so not looking at it further.

Literally spent a couple of days on it. My motivation is also not rays but elliptical cones (for wave optics). A BVH does significantly better than a kdtree for cones, so I also put together an 8w BVH.

Yes. I’d also assume it’d be different on a GPU.

Exactly what I found actually. Compressing the nodes reduces performance even with moderately complex scenes. Uncompressed 8w nodes is fastest.

Hit me up if you are in Tokyo!