INSIDE THE DLSS 3 TECH

APC’s Bennett Ring asked Nvidia to explain more about DLSS 3, and what the future holds.

Does the resolution upscaling within DLSS 3 differ from DLSS 2 in any way?

No. Super Resolution and Frame Generation are independent options included in DLSS 3. DLSS Super Resolution has improved significantly over the years in both image quality and FPS, and we anticipate continued improvements to DLSS Super Resolution going forward.

Why the confusing naming of DLSS 3, as DLSS stands for Deep Learning Super Sampling, yet DLSS 3’s feature has nothing to do with super sampling?

DLSS is a pioneering AI-powered technology that improves on all three aspects of gaming graphics rendering: image quality, frame smoothness, and low latency responsiveness. This is true for DLSS Super Resolution, and it is also true for DLSS Frame Generation. We felt that since the goal was the same, it made sense to refer to these technologies together.

One of the few noticeable visual issues with frame generation occurs when scenes change, or games swap from a static menu screen to in-game. Will frame generation technology be improved to remove this issue? Is it possible to remove the HUD from the area being treated by frame generation, as we’ve seen with DLSS 2?

The scene change issue is considered a bug and should be fixed. Games already provide a flag to TAA or DLSS in order to deal with scene changes, but this flag was not properly handled in DLSS 3.0. This will be fixed quickly. We also have some changes to how DLSS Frame Generation handles the HUD that we expect will significantly improve HUD generation quality in the short term.

Does Nvidia recommend frame generation for games with lower frame rates (4060fps), or is it best suited to games which already have high frame rates (100fps+)?

To achieve the best gaming experience with DLSS Frame Generation, the AI network’s input frame rate must still perform at a level such that the main game loop, input handling, and simulations run well. If these elements aren’t updated fast enough, the game itself can feel sluggish.

The exact input frame rate needed varies based on game genre, the game engine and other factors. DLSS 3 includes DLSS Super Resolution which helps boost the input frame rate (including the main game loop) for DLSS Frame Generation by up to 2x. For example, the game could render natively at 25-30 FPS, be boosted by Super Resolution to 40-60 FPS, which provides a solid input frame rate for the Frame Generation AI network to boost FPS by up to another 2x (to 80-120 FPS).

Note that RTX 40 Series GPUs have the raw performance to render the majority of games well above 25 FPS at native resolution. Particularly as we address some of the UI image quality issues mentioned above that become more apparent at lower FPS, we believe input frame rate concerns will become even less of a factor.

Currently V-sync isn’t officially supported with frame generation – will this change in the future?

Several press also provided feedback that they would like V-sync to be supported with DLSS 3. We agree and have improvements for V-sync support in the works.

How long does it take in man-hours for a game developer to implement DLSS 3?

DLSS Frame Generation is implemented using Nvidia Streamline, and it requires essentially the same inputs as DLSS Super Resolution, with the additional requirement of implementing Nvidia Reflex to reduce latency. We learned a lot about how to integrate DLSS into games over the past four years and we build upon that experience to making the DLSS Frame Generation as easy to integrate into as many games as possible.