I personally know some of the guys who developed the CNN based denoising. This stuff really works and it is actually pretty simple to implement compared to the

I personally know some of the guys who developed the CNN based denoising. This stuff really works and it is actually pretty simple to implement compared to the truly horribly complicated hacks that you need to get e.g. realtime reflections even remotely to a level that is not outright jarring and obviously wrong looking. What has me worried is not the performance of the ray-scene intersection, but the dynamic BVH update that is required for each frame. Building acceleration structures for ray intersection tests is a hard tradeoff between fast construction times and fast intersection tests. With animated/deforming objects, this is quite an interesting challenge. But we might eventually see some dedicated hardware for that, too. Academic implementations exist.