NVIDIA DLSS 3.5 was announced at this year’s GamesCom 2023 opening night and is said to arrive this coming fall. The next iteration appears to have a focus on a new feature called Ray Reconstruction (RR) to provide significantly improved visuals when using ray tracing effects. RR is a type of denoiser and is said to have been trained on five times more data than DLSS 3. According to VideoCardz denoisers often create graphical anomalies due to a lack of data, but also strip away data that is needed for upscaling while also reducing color data.
Per VideoCardz:
“The utilization of denoisers was found to strip away essential data required for upscaling, leading to a loss of color data during denoising and subsequent upscaling. Furthermore, traditional methods may result in inaccurate lighting effects due to the accumulation of pixels from previous frames, often resulting in ghosting.”
“Denoisers also contribute to subpar global illumination and lower-quality reflections, as there isn’t enough data for interpolation. This will not be the case with Ray Reconstruction, which is the primary feature of DLSS 3.5 update.”
How DLSS 3.5 Works
According to slides provided by NVIDIA, Ray Reconstruction performs through four essential steps. RR is first trained to incorporate addition engine and software data, recognize the ray-traced effects, analyze and separate the good and bad temporal and spatial pixels, and then ultimately retain the high-frequency data needed for improved upscaling. Ray Reconstruction can also provide additional FPS in gaming over previous DLSS technology when combined with super-resolution (DLSS 2) and frame generation (DLSS 3).
NVIDIA DLSS 3.5 is set to arrive this fall and will debut in CyberPunk 2077 Phantom Liberty (September 26), Alan Wake II (October 27), and NVDIA’s developer tools, Ominiverse Platform, Chaos Vantage, and D5 Renderer.
Discussion (19 replies)
Join Discussion →This is massive. Up to like a 10% fps increase using it with FG and DLSS in cyberpunk from the example they showed.
I know very few games will get the tech and even fewer will get as much of a benefit but I'm excited.
The point isn't the performance increase, the point is the image quality improvement vs a traditional denoiser. RR also incurs a cost, just like a denoiser, but the point is AI can do it better, that's the claim anyway.
Hopefully this will help with RT 'swimming/noise' especially on lower resolution DLSS.
If it works as they advertise haha
I think the intended benefit is when RT is used with upscaling, it tends to lose quality due to temporal AA, the denoiser has artifacts, but with this new AI method it employs a different method to find patterns to fill in the empty pixels that can do a better job at it when using upscaling with temporal AA.
I wonder if this will eliminate the trails of moving objects off in the distance. It's not as bad as it used to be but it is still there with DLSS 3.
So the tensor cores will do the denoising vía AI, instead of just using a denoiser filter algorithm.
When rtx was first announced, I wandered why did it use an algorithm instead of AI, I guess it wasn't ready just yet
Interesting that it works on previous gen cards as well. Not that I'm complaining just uncharacteristic.
Previous gen cards that support DLSS3. so not really.
[URL unfurl="true"]https://www.theverge.com/2023/8/22/23841148/nvidia-dlss-3-5-ray-reconstruction-ray-tracing-quality[/URL]
?♂️
Separate Frame Generation and DLSS3 in your mind. DLSS 3 works on all RTX GPUs, RR works on all RTX GPUs, Frame Generation only works on RTX 40 GPUs.
Nvidia had to have made this mess on purpose, this is some AMD PR-grade confusion...
Yea clearly it's confused the eff out of my old ass.
Yeah, it's become a cluster f in trying to decipher. Happy for the features but it is getting confusing and game dev's are not even consistent if/where they put the toggles in. Some are under display and some under graphical features. For most of us that doesn't really matter since we're going to hunt and tweak wherever anything is anyway but I have noticed that.
They showed off this in the Alan Wake 2 trailer so it is massively improving the IQ and response time in 'fully raytraced' scenarios
One of my big complaints with PT in CP2077 was the smeariness of the fences and barbed wire.
If I'm not mistaken DLSS3=DLSS2+Framegeneration (only on RTX40), but DLSS3.5 = DLSS2+AI Denoising on pre RTX40 but also = DLSS2+Framegeneration+AI Denoising on RTX40 (and up?).
There have been rumors that frame generation could be coming to RTX30, but I don't think so.