@nilsonlim DLSS is just a realtime tree lookup of a neural network index for upscaling. It doesn’t use all 9 Tflops of the Tensor cores to run the DLSS lookup, and plenty of things Tensor cores are good at run almost as well in CUDA. The only requirement is that the source frame resolution matches the training data, which up until now has been 1440p as the reference and possibly up to 16k as the training data.
You could conceivably train some high res RTX results against some low res RTX results with a particular feature set to get 4k RTX gaming on a 1080, just not with a source of 1440p. So all the data would need to be retrained. Maybe they could target just the RTX combined results in low res + DLSS and combine it with a high res diffuse scanline layer. Maybe 720p. I’m sure there’s a balance that would let you target 60fps.
Unfortunately with the high cost of training DLSS at the moment we’d need NVidia’s cloud DLSS training service available to try it. But I’m sure it’s possible and we could even do it in Unreal Engine. The other neat thing about only training your RTX combined buffers is the single set of result might be usable on *any *game using that method.
It’s worth noting that the AI denoiser is also an implementation of the tech behind DLSS, and you’ve seen how much that can clean up your image.