Biggest win here is that if this works as promised, Nvidia will soon find "a way" to bring DLSS 3.X to other series of cards beside the 4XXX.
Quoting myself from a previous thread
“Peoples can validate that for themselves if they think old RTX cards are left out. You can test by using either the older Video Code SDK that provides optical flow analysis functions for hardware as old as Maxwell or the new Optical Flow SDK for Turing>.
Prior to Ada the way of loading data into the OFA was either from NVDEC or from very slow managed memory buffers and even tho D3D buffers were supported it the latency was nowhere near useable for interactive real time rendering.
With Ada the OFA fixed function units can access the entire memory directly via address masking, as well as the L3 cache and likely have had a substantial number of other hardware specific changes to facilitate this. Ada’s higher clock speed also helps.
Ampere is actually slower than Turing for motion vector extraction, but has fewer features such as not going under 4x4 grid, while FG uses 1x1 or 2x2
At same clock to Ampere, Ada is 2.5 to 4 times faster than Ampere. Leave the clock of Ada back to normal and it’s even higher.”
I would wait to see the latencies of AMD side before jumping to conclusions that it’ll be worth it. Especially since DF says that it’s compute dependent.. the kind of games that NEED frame gen to be playable are typically compute heavy, we’re talking Cyberpunk 2077 overdrive kind of “need”. Otherwise, if your goal is to go from 120 fps to >200 fps, i would but a big question mark on latency. AMD solution seems akin to the Turing optical flow.