I believe DLSS frame gen predicts future frames (X_1, X_2, X_3) given (X_0, X_-1, X_-2, ...), without waiting for X_4. At least that's the impression I get from their marketing.
Yeah, but there's still a latency penalty, because X_1, X_2, X_3 won't respond to player input, so your effective latency is still that of your 'real' FPS, and that's lower than without because the frame gen takes a good fraction of GPU resources.
Nvidia Reflex 2 does that with async frame warp (that has been used in VR for a while now), but it's separate from DLSS, and is not supported in many games.
