Custom graphics architectures aren't always a disaster - the Switch 2 is putting up impressive results with their in-house DLSS acceleration.
Now, shackling yourself to AMD and expecting a miracle... that I cannot say is a good idea. Maybe Cerny has seen something we haven't, who knows.
The Switch 1 also had CUDA cores and other basic hardware accelerators. To my knowledge (and I could be wrong), none of the APIs that Nintendo exposed even gave access to those fancy features. It should just be calls to NVN, which can be compiled into Vulkan the same way DXVK translates DirectX calls.
It's better off if I let Digital Foundry take it from here: https://youtu.be/BDvf1gsMgmY
TL:DW - it's not quite the full-fat CNN model but it's also not a uselessly pared-back upscaler. Seems to handle antialiasing and simple upscale well at super low TDPs (<10w).
It's literally not. From the description of TFA:
In this video, Alex goes in-depth on Switch 2 DLSS, confirming that there are actually two different forms of the technology available - the DLSS we know from PC gaming and a faster, far more simplified version.
The entire Switch 1 game library is free to play on emulators. They probably put a custom accelerator to prevent reverse engineering. A consequence of using weaker spec parts than their competitors.