what did I just read ? rt stands for ray tracing.
dlss works on tensor cores (fp16/8 accelerators), and they are different indeed, cause newer generations have half of old gen tensor core number per sm but they're twice as fast.
2060 Super has the same number of tensor cores as 3080 and more than 4070Ti.
You think that implementing dlss is a separate process for each generation, which is just wrong. Each new generation of tensor cores serves the same purpose and requires no individual tuning, it just saves on die space by making them more efficient.
the only cards that will need optimizing for different versions of same upscaler will be rdna2 when future versions of fsr use rdna 3's ai hardware for training the algorithm, as the spatial upscaler in fsr2.x probably hit a wall in what it can do.
it was kinda short sighted of amd not to put those units on earlier rdna (2/1) as they're clearly planning on using them for something, but the development of a true image reconstruction technique will be hindered by backwards compability.
same will be true about fsr3, which runs on async compute now, while on nvidia it runs on beefed up OFA on rtx40.I don't have a problem with nvidia saying "our cards from rtx40 onwards will have better OFA for doing FG", I do have a big problem with counting FG as part of generational leap, but that's another issue, unrelated to the technical side of how frame generation is handled on nvidia and amd. Using async compute for frame gen already had amd admitting the results will vary during fsr3 presentation, namely the more a game uses async compute already, the slower it will run. It's not the only limitation of fsr3, as not using OFA will provide less information to the gpu too, to the extent where fluid motion frames tech will be disabled when mouse movement is too fast.
frs3 is intended to run on framerates that are high already (+60 at least), while I've seen people saying that nvidia's frame gen does quite well at 30-40fps already, just not 20s as the lag will be awful even with reflex on, and there will be gameplay-breaking visual inconsistency between frames.
To sum up, fsr3 needs 1. async compute headroom 2. high fps to run well, while dlss3 doesn't really care that much, as long as the framerate is above 30.
Sorry for going a bit OT, but I hope fsr3 vs dlss3 will need no further clarification as far as their technical side goes.
I think it's not free in starfield only, and while he did that, another dude made one free for anyone to use.