| JRPGfan said:
The DLSS type it is running seems to be a differnt model than the ones Nvidia uses normally. They think its a "super light" model, and the low resolutions and motion blurr, hid it from them. (ei. the quality was so low, they didn't think it used DLSS at first).
|
Just want to point out that this isn't exactly what they said. A "legendary super light model" based on "Nintendo's patents" was one of the speculated possibilities posed in the question, but not something speculated about in the analysis much at all.
The conclusion actually made, was that because the post-processing effects are done at sub-native resolution and introduce much more dominant aliasing, it was hard for them to notice DLSS-specific aliasing patterns. They never said they think there is evidence for "a super light model".
A lot of this is also based on the assumption that the game uses a CNN model (or the Switch 2 is incapable of using ViT's) and the artifacts from that model are going to show up as such. But ViTs are scalable by parameter count just like CNN models, and there is no reason to believe that they couldn't have trained (or, more likely, quantized) a lower-parameter ViT model than what Nvidia has for PC's which aren't over-utilized by the tensor cores and give better output. Heck, it still isn't clear that the Switch 2's tensor cores are even a bottleneck. Alex did find the sort of CNN model artifact he was looking for, but that could just be confirmation bias now that he knows that DLSS is there, in fact, or it could be something that shows up in a lower-parameter ViT model.
All of this is to say that they are going to have to see more Switch 2 implementations of DLSS to be able to reliably identify it in Switch 2 games.
Edit: I train ML and DL models for a living. Nvidia almost certainly has dozens to hundreds of trained model revisions for each one they actually release, at different parameter counts and which have different artifacts in their output. They probably do this for both CNN and ViT's (as well as hybrid compact transformers.)
For PC they pick one that targets the minimum expected hardware, say an RTX 2050, but they could pick something different that is better suited for Switch 2 as a platform if the tensor cores are over-utilized. We've already seen that they have tailored different ViT models to have different preferred artifacts.
Last edited by sc94597 - on 20 April 2025