Thanks a lot for the input.
Ok, I was even ready to cook before and deploy the whole terrain volume(at game installation) as a single reference SVO(even if very heavy on disk), stream / play with / destroy an instance of it in-game.
Use this to generate your not-yet-destroyed object once at load time, not for real time.
Upon impact with shots and let the gravity break-up the detached group(s) in smaller shards upon collision(+enough speed) with the underlying rest of the terrain.
Depends on how you want to destroy it.
If I use child pointers, I am not sure to understand how to update the octree after such events.
Thanks for that, this is one of the keys I needed to hear !
To shatter an SVO is just taking a chunk of the voxel nodes out and creating a new SVO out of them.
I tried to get familiar with several inner/leaf node structures to store my noise values but I am still hesitating in choosing one that will serve as well an easy ray-casting and physics animation.
Heavy use of spatial partitioning
I saw one person on youtube publish a sample with fair frame rates, ray-casting only on the CPU.
I asked him about the node structure used and he said he used barely half of Laine & Karas's one (without contours nor colors for test purposes) and their exact same ray-casting algo but without CUDA. Laine's ray-caster looks indeed easy to pull from CUDA to CPU. I probably need to have a shot at something like this but I am still not 100% sure :
Much thanks again Digitalfragment.