NVIDIA’s FSR 3.0: Igor’s Lab claims (opens in new tab) that Nvidia’s DLSS 3.0 frame generation is compatible with upscaling techniques other than DLSS. Through extensive testing, Igor was able to utilize Nvidia’s frame generation technology in tandem with AMD’s FSR 2.1 and Intel’s XeSS temporal upscale to play Spider-Man Remastered.
So now we know what to expect from AMD’s forthcoming FSR 3.0 release, and we can also speculate on whether or not Intel may integrate frame generation into XeSS in the future.
Nvidia’s newest iteration of its Deep Learning Super Sampling (DLSS) algorithm is version 3.0. Even still, contrary to expectations, version 3.0 does not improve the super sampling part of the technology but rather includes AI frame generation. They make it possible for Nvidia GPUs with support to achieve far higher frame rates than would be possible at native resolution.
When the technology is implemented in a game, it works by adding a single artificial intelligence-generated frame between two traditionally displayed 3D frames. However, a drawback of this kind of frame creation is an increase in input lag, which is the time it takes for on-screen actions like clicking or pressing a button to be reflected in the game.
As a result, Nvidia requires all DLSS 3.0-supported games to include its Reflex technology in order to bring input latency down to a playable level.
Nvidia’s DLSS 3.0 is notable for the fact that it splits the frame-generating process from the upscaling process. This means you may use the frame generator at its native resolution or in conjunction with other upscale, as the two functionalities are completely independent of one another.
Igor’s testing relied on this feature, and he confirmed that this is how Nvidia designed DLSS 3.0 to function, thus it’s not a problem.
Benchmarks
By running Spider-Man Remastered on a GeForce RTX 4090 GPU with all of the upscaling settings enabled, Igor compared the frame generation of DLSS 3.0 with Nvidia’s true DLSS upscaler, FSR 2.1, and XeSS.
The average number of frames per second the 4090 could generate at its native resolution was 167.6 when frame generation was turned on. Intel’s XeSS Ultra Quality mode, the closest and most demanding upscaling mode, is just 1 frame per second faster than native.
XeSS balanced is at 192.7 FPS, while XeSS Quality Mode is at 183 FPS. With DLSS Quality at 197.4 FPS and FSR Quality at 193.8 FPS, both are faster than the XeSS balanced mode. The next three modes in order of FPS are XeSS Performance (203.9), FSR Balanced (206.8), and DLSS Balanced (209.0).
Since Intel’s XeSS upscaler doesn’t include an ultra-performance mode, the performance chart is topped off by FSR and DLSS modes and ultra-performance modes. Therefore, the results for DLSS performance and FSR performance are virtually identical, at 220.1 and 220.5FPS, respectively.
While FSR Ultra Performance offers the highest frame rate (231.4 FPS), DLSS Ultra Performance is the quickest setting (224.0 FPS).
In sum, this suggests that DLSS and FSR 2.1 work similarly, with little discernible frame rate differences in actual gameplay. However, XeSS looks to be more demanding, with its fastest settings being a whole rung slower than the somewhat higher-quality models of FSR and DLSS.
Then there’s XeSS Ultra Quality, which is the slowest option by a wide margin, even being slower than the native resolution by a single frame per second. However, this can still be useful, as XeSS’s anti-aliasing could be superior to that of the game’s native TAA, SMAA, or other AA alternatives.
Comparison of Generated Frames With Natively Rendered Frames
Any upscaling mode falls well short of the native frame rendering capabilities of the RTX 4090. It’s clear that the CPU is the limiting factor here, as the game maxes out at 133–135 FPS across practically all settings, with the exception of native resolution, FSR Quality, and Ultra Quality, which all run at around 125–130 FPS.
Yet even when comparing these three slower modes, the frame rate differences are so small that they are essentially imperceptible.
Contents