Worth adding that if we get more hardware support for deep neural networks on the GPU (e.g. tensor cores), training a denoising DNN on the noisy renders can save a lot of computation and fits neatly into the pipeline.

If the ray tracing can be done at or around 1 ray per pixel or less using a network trained to do merge information over multiple frames and upscale we could probably get away with less. Maybe more if we can feed in a depth map, velocity map and a flat shader less rendering or other information to help guide the DNN.

Might even end up faster than current raster renders.

See https://www.chaosgroup.com/blog/experiments-with-v-ray-next-...