A process that should also be used by AMD’s FSR 3.0.
DLSS 3.0 (Deep Learning Super Sampling) uses a mechanism of Frame Generation which generates new images via AI. As always, the goal is to increase the number of frames per second, including when performance is constrained by the CPU. Our colleagues from Igor’s LAB show that this technology is not limited to DLSS but that it works with other upscaling solutions, namely AMD’s FSR and Intel’s XeSS.
First of all, a few explanations on the Frame Generation. To summarize very briefly, DLSS 3.0 couples the super-sampling part with the generation of AI images, by adding an AI-generated image between two “classic” 3D images. A good way to increase overall performance, but comes at the cost of higher input latency. NVIDIA counters this phenomenon with its Reflex technology, which is now mandatory for DLSS 3.0 games.
Optical Multi Frame Generation
Here are more details from NVIDIA:
“The Optical Multi Frame Generation generates entirely new images, rather than just pixels, which provides significant performance gains. The new optical flow accelerator built into the NVIDIA Ada Lovelace architecture analyzes two sequential in-game images and calculates motion vector data for objects and elements that appear in the image but are not modeled by the traditional game engine motion vectors. This greatly reduces visual anomalies when the AI renders things like particles, reflections, shadows, and lighting.
Pairs of super-resolution game frames, along with the engine and optical flow motion vectors, are then fed into a convolutional neural network that analyzes the data and automatically generates an additional frame for each frame rendered by the game. – a first for real-time game rendering […].
Since DLSS frame generation runs as a post-process on the GPU, it can improve frame rates even when gaming is CPU limited.”
What is interesting, and which allows the experimentation reported here, is that within DLSS 3.0, NVIDIA distinguishes the component Frame Generation and the scaling component.
The results
Our colleagues tested the GeForce RTX 4090 and a Core i9-12900K at 2160p with the max settings. As you will see, the FG boosts the number of frames per second with DLSS but also with Intel XeSS and AMD FSR. On the other hand, the FG needs to be used in conjunction with one of these technologies to limit variations.
Our colleagues conclude: “NVIDIA’s FG implementation works with DLSS, but also with the FSR and XeSS of the two competitors. There is almost no difference between DLSS and FSR when it comes to performance, only XeSS lags a bit. […]”. They believe that “if with FSR 3.0 AMD manages to achieve such a good implementation for images generated through it and also offers it to older graphics cards or even competitor cards”, the company will do a very good job. operation.
Sources: Igor’s LAB, NVIDIA