By using this site, you agree to our Privacy Policy and our Terms of Use. Close
sc94597 said:
HoloDust said:

Yes, Switch 2 with its meager Tensor cores is no doubt doing better than nVidia's PC GPUs...

/Picard-Riker-Worf triple facepalm

To be fair, there are odd things about Switch 2's DLSS implementation and its performance costs.

For example, based on their simulated testing of DLSS latency on an RTX 2050 Digital Foundry thought upscaling to 4k 60fps would be impossible due to the frametime penalty (18.6 ms), but it looks like Fast Fusion is doing that. (Although DF's simulation was probably flawed from the start, in its assumptions.)

https://youtu.be/JJUn7Kc3W3A?si=3naNeN2QiM5T_gwf

When Nvidia trains these models, they probably are training dozens to hundreds, most of which they reject. It is possible that one of those rejected models or a distillation of a non-rejected model, or even a brand new model is what is running on Switch 2. Also a lot might be achieved through pruning a model when targeting a given input range, which is far narrower on Switch 2 than PC. Basically you might be able to reduce parameter count (and therefore inference compute) without reducing quality too much  because certain nodes might only activate when the input resolution is higher than what Switch 2 typically achieves. Pruning can improve the model's efficiency. Although this begs the question of why Nvidia hasn't tiered DLSS like this on PC for different hardware ranges. Development complexity might be part of why. Although it might just be the case that there isnt much to prune. 

Personally, having played a few games (Cyberpunk, Hogwart's Legacy) I don't think the Switch 2's DLSS implementations are much better or worse than what we've seen on PC performance modes @ 1080p. Even if the tell-tale artifacts are hard to identify, it could be just that we are making wrong assumptions about the models being used. We still don't know what sort of model or models are even running on Switch 2. The use of DRS and unique post-processing techniques make this even trickier. 

I guess we'll see the results in time - IIRC (I watched that 2 days ago) DF specifically called out DLSS disocclusions in Fast Fusion and problems with image quality, so waiting to see how Hogwarts fairs in their tests, since that was worst case scenario with edges falling apart in quite a few places in that video they were commenting on some time ago.