Also, correct me if I'm wrong, but DLSS is done on hardware (can only be done on nVidia RTX GPUs), whereas FSR can be done on modern GPUs (based on RDNA 1 and newer) in general. DLSS being done on hardware also, I believe, gives DLSS some advantages over FSR (quicker processing)
DLSS
is accelerated by some extra hardware, and it does give it an advantage, but that's primarily in quality. Speed is weirder?
From a quality standpoint - you don't actually want the most accurate upscaler. Well, you
would if it was 100% accurate, if it gave you the exact same thing as rendering at higher resolution. But there will always be errors. Which errors matter more to the human brain, unfortunately, are highly situational.
FSR and DLSS work the same way. They take pixels from previous frames, and combine them to add detail to new frames. They get extra information from the game engine to inform those decisions. However FSR uses a bunch of hand written rules for the combination process, DLSS uses AI.
FSR is pretty good at finding detail from the old frames and putting it in the new frame. Where it falls apart is making sure those new pixels add up to something that makes sense to the human eye.
DLSS uses AI. Since AI works statistically, instead of having a bunch of human-written rules, AI is better at capturing "okay, you know what, this is
probably a situation where we should connect these pixels into a line." Or "we kept this pixel last time, this looks like a situation where we should keep it because blipping it in and out of existence will be more distracting, even if it's slightly inaccurate."
DLSS's AI uses tensor cores, FSR's human written code runs on the regular GPU. In general, these two things tend to run about the same, not for technical reasons but practical ones - both of them need to run on similar spec hardware, in a similar amount of frame time, and so they are optimized accordingly.
Where speed is different is in how the algorithm works. DLSS needs to make a decision for every pixel in the final image, regardless of how much data goes in. That's why 4k DLSS times are the same, no matter the quality mode. FSR makes a decision for every
new pixel the upscaler is adding. That makes FSR speed depend on the quality mode more than the output resolution.
DLSS runs on hardware that can run concurrently with rendering, FSR doesn't. That let's DLSS offer a special parallel upscaling mode where you upscale one frame while rendering another. To my knowledge no game implements that on PC, but it's potentially very interesting for consoles