Actual paper here: http://cg.ivd.kit.edu/svgf.php
No neural networks needed, successful with just 1-spp.
From what I can tell from the papers, the temporal-denoising was the big advancement. Simple spacial denoising has been done for a while, but the "denoised blobs" kinda flow around in an animation (the "clusters" of shadows and/or fireflies move randomly, as per the Monte-carlo raytracing algorithm demands). By denoising over time, you end up with not only more effective samples (smoothing over 10-frames means you really have 10-samples you've averaged over), you also end up with a smoother animation without "floating blobs".
That's fantastic! Mostly-temporal techniques wouldn't do the trick either, but they've got a mix working. 1080p40-50 on a Titan Xp can't be mainstream for a while (they quote a similar pixel rate for the actual tracing as for their filtering), but there'll probably be quite a few iterations on it over the next few years.
With all that specialized hardware, Nvidia gets some interesting choices of how to arrange work. Hopefully latency is unchanged.
Navi, the PS5, and the Xbox Two will be interesting. If Navi has any particular features at all to help out raytracing it's probably going mainstream the moment the PS5 is out, and if not there will be a lot of incentive to make it work anyway. After reading that paper, making it work despite no special support seems a lot more feasible.
While this paper / youtube video doesn't need any neural networks for the image... I bet you that it uses those FP16 Tensor Cores (super high-speed low-precision arithmetic).
It looks to me like it's probably memory-bound.
IIRC, NVidia was hoping to test 0.5 or even fewer samples per pixel. Because the time-averaging effect really does work some magic. 0.5 spp effectively becomes "denoise an average every 2 frames" for example. Moving objects would probably be an issue, but being able to just do 0.5 or even 0.25 spp on all static objects (then doing a more expensive pass on dynamic objects) probably would save a lot of time.
At some point (maybe even immediately), it's going to be better to do all tracing-related work at a reduced resolution instead of spending that much work filtering a small number of samples. At least at 1 spp, the most obvious difference from reference looks to just be the blurriness anyway. 1 spp at 720p upsampled onto a normal 1440p render might both look better (overall) and run faster than the 1 spp 1080p demonstrated here.