DLSS 5 - Generative AI

erek said:
excited?
Click to expand...

Very. Although I must admit, after gaming at 1440p for the past month, it really isn't that terrible anymore thanks to DLAA. 1440p with anything else whether it's FXAA, SMAA, and especially TAA (🤮) just looks so disgustingly blurry/jagged. Still, more pixels is more pixels and can't go wrong with that.
 
The new version of DLSS5 (3.9.16) that dropped today gave me a 10FPS boost to my base frame rate in cyberpunk at 4k. So up to 60FPS if you are using 6x frame gen. I'm using 2x frame gen so I got a 20FPS boost overall. Pretty happy with that. I'm still using preset B for DLSS because it looks much better than preset C in Cyberpunk and preset C has some micro stutters on my system while preset B is flawless. Preset C does perform about 10FPS better than preset B though.
 
Comixbooks said:
https://www.pcgamer.com/hardware/gr...eeping-it-the-hell-away-from-game-characters/

About the performance update using less resources with the new version.
Click to expand...
“A new version of DLSS 5 on NBA 2K27was just released, and modders have already extracted it and are running it on unsupported games. Modders are comparing performance between the older and newer Neural Rendering (NR) model and found it up to 40% faster than the previous one.

Official DLSS 5 running on unsupported games delivers up to 90% better performance than the leaked version — Nvidia says the new NR model uses 20% less VRAM and processes neural workloads up to 2x faster
​

Nvidia's new Neural Rendering model is smaller but better.”

https://www.tomshardware.com/pc-com...nd-processes-neural-workloads-up-to-2x-faster
 
In this video, Hardware Unboxed tests Nvidia's new DLSS 5 neural rendering model (released in October), which claims to run twice as fast as the original version while reducing VRAM usage by 20% (0:00 - 0:30). The host evaluates these performance claims using NBA 2K27 across the full GeForce RTX 50 series GPU lineup.

Key takeaways:

• Performance Gains: The new model delivers significant speed improvements—averaging about 83% faster at 4K and 94% faster at 1440p across the RTX 50 series (9:42 - 10:06).
• Usability: While the performance hit is still substantial compared to disabling the feature entirely, the new model makes DLSS 5 much more viable for mid-range cards like the RTX 5070 and RTX 5060 at 1440p (12:42 - 13:04).
• Visual Quality: The host finds that image quality remains largely unchanged and that the original visual artifacts—such as lighting flicker, inconsistent skin tones, and halo effects—persist in the updated model (14:30 - 16:59).
• Conclusion: While the October update is a necessary step in the right direction, the technology still requires further optimization in both performance and visual stability before it can be recommended for widespread, everyday use (18:10 - 18:48).”


View: https://youtu.be/k9MQa9pdj8Q?is=05qCH4Ve2XZOyjKk
 
Last edited:
The new model definitely runs twice as fast, but your actual performance gain will depend. Here's an example using CP2077 at 1440p.

OLD Model (Preset A):

1792448745776.png


NEW Model (Preset B):

1792448765849.png


So using the old model, my fps in CP2077 is 78 which is a render time of 12.8ms. Out of that 12.8ms, 3.6ms was used for NR at 1440p resolution. The new model runs twice as fast so instead of a 3.6ms render cost, that is now just 1.8ms. 12.8 - 1.8 = 11ms is now the total render time with the faster model and so the new fps should be 91, and it is so the math completely checks out. In this scenario though, the new model only translates to an actual performance gain of just 16% over the old model, but the gains at 4K will be far more significant. Power consumption did drop by 21 watts at 1440p so there's probably an even bigger power savings at 4K.
 
Last edited:
MistaSparkul said:
The new model definitely runs twice as fast, but your actual performance gain will depend. Here's an example using CP2077 at 1440p.

OLD Model (Preset A):

View attachment 832108

NEW Model (Preset B):

View attachment 832109

So using the old model, my fps in CP2077 is 78 which is a render time of 12.8ms. Out of that 12.8ms, 3.6ms was used for NR at 1440p resolution. The new model runs twice as fast so instead of a 3.6ms render cost, that is now just 1.8ms. 12.8 - 1.8 = 11ms is now the total render time with the faster model and so the new fps should be 91, and it is so the math completely checks out. In this scenario though, the new model only translates to an actual performance gain of just 16% over the old model, but the gains at 4K will be far more significant. Power consumption did drop by 21 watts at 1440p so there's probably an even bigger power savings at 4K.
Click to expand...
kram182 GoldenTiger

​

“NVIDIA's Faster DLSS 5 Model Gets Extracted From NBA 2K27, Modders Report Up to 90% Gains

NVIDIA shipped its faster DLSS 5 Neural Rendering model with NBA 2K27 on October 16, and it didn't take modders long to pull it out and test it elsewhere. Renan Maniero on X extracted the new 310.9.16.0 build of the DLL. That lets people inject it into games that don't officially support DLSS 5 through tools like RenoDX and OptiScaler. The new DLL also keeps the older model around as Preset A, with the new one sitting under Preset B, so comparing the two is as simple as switching presets. Gaming Optimized on X did exactly that, and found the new model running up to 90% faster than the old one in The Witcher 3. The channel points out that the gain varies from game to game, so 90% looks like the best case rather than the norm.

That roughly aligns with NVIDIA's own claims. The company says the October update shrinks the model by 24% to roughly 118 million parameters, cuts VRAM use at 4K by 20%, and runs the neural rendering workload up to 2x faster than the September version. NVIDIA also quoted up to 40% higher frame rates in NBA 2K27 specifically. Officially, DLSS 5 is still limited to RTX 50-series GPUs and GeForce NOW. NVIDIA plans to bring it to the RTX 40-series, though there's no date yet, and there's no word on RTX 30 or RTX 20 support. Modders, apparently, aren't waiting, since the OptiScaler README notes a modified runtime that gets it running on RTX 20-series cards as well.”
 
Last edited:
I'm not going to make claims about general operation across all brands of cards but for me. My power usage was around 590 to 614 watts. With the latest driver from Nvidia it dropped to 590 to 605. With the latest update to DLSS 5 it's now sitting at 565 watts all the time. This is in Cyberpunk at 4k with Reno DX DLSS 5 profile 1. DLSS 4.5 Ray reconstruction profile F and DLSS profile C with path tracing, Ultra plus, HDR and about 430 mods with an overclocked card and no under volt. I can't imagine a game that can push the system harder than this. I would say the power usage is no longer being pushed past the safety limits even under the harshest conditions, for me at least.
 
  • Like
Reactions: erek
like this
MistaSparkul said:
https://wccftech.com/witchfire-crea...ls-it-amazing-expects-dlss-3-acceptance-path/

View attachment 832242

The haters have definitely gone pretty silent now compared to when DLSS5 was first announced :D . Even John from DF who acted like he hated it with every fiber of his being is now starting to accept that it actually looks good when used properly.
Click to expand...

Let’s just suppose: Nvidia is doing with DLSS 5 what it did with ray tracing—reducing the data processed while making the difference barely noticeable. This way, the FPS drop is minimized, and every zealot is satisfied.
But did they actually improve something to make lighting more realistic? Or is it still just a visual trick?
 
AVATARAT said:
Let’s just suppose: Nvidia is doing with DLSS 5 what it did with ray tracing—reducing the data processed while making the difference barely noticeable. This way, the FPS drop is minimized, and every zealot is satisfied.
But did they actually improve something to make lighting more realistic? Or is it still just a visual trick?
Click to expand...

Dude, these are video game graphics, it has ALWAYS been about visual tricks. If this the end result of such tricks then I couldn't care less. This zealot will continue to enjoy the DLSS5 slop.

1792605849287.jpeg
 
AVATARAT said:
Let’s just suppose: Nvidia is doing with DLSS 5 what it did with ray tracing—reducing the data processed while making the difference barely noticeable. This way, the FPS drop is minimized, and every zealot is satisfied.
But did they actually improve something to make lighting more realistic? Or is it still just a visual trick?
Click to expand...
All aspects of video game development are "tricks" or we'd just be watching lines of code scrolling across the screen.
 
As mentioned earlier, DLSS5 uses some elements of RT, but its actual purpose is to make the image look nice without paying much attention to where the light is coming from.
 
AVATARAT said:
As mentioned earlier, DLSS5 uses some elements of RT, but its actual purpose is to make the image look nice without paying much attention to where the light is coming from.
Click to expand...
I think the RT-lighting noisy passes are one of the DLSS5 input it can receive too, that why it does not have to pay much attention because it has the game raytracing engine that did it already, like for DLSS RR api, the result of those rays and what those light position/bounce mean has been taken into account, I would imagine most//all of those are being used:

https://github.com/NVIDIA-RTX/Streamline/blob/main/docs/ProgrammingGuideDLSS_RR.md

This is by design to work with RT not to replace it, a super denoiser for very limited pathtracing it if you will, filling the gap that you always need to fill because of real time limitation with more knowledge on what it would have probably been with 1,000 rays per pixel than before, that why we see the final output change in games when they add more RT even if DLSS5 is on.
 
Yes, it uses RT as a base (when available), but ultimately it's the AI filter that makes things look more attractive – that's by design.
 
AVATARAT said:
Yes, it uses RT as a base (when available), but ultimately it's the AI filter that makes things look more attractive – that's by design.
Click to expand...
I am not sure of the point trying to be made, yes the neural rendering model to estimate what the denoising of the pathtracing should look like, subsurface and material details inference (guessing) is what make things look more attractive, filter is a wrong word for this, filter are purely destructive not additive like this is.

Filtering is usually substracting-denoising high frequency details (think anti aliasing, remove noise by adding blur), this is a additive/generative step, one that do take a lot of the lights situation in place doing so via the ray tracing results as an input (we can see big shift when LOD decision remove things from the raytracing and the info DLSS5 get with the camera getting closer or father because of this for example).

What's mostly going here is removing the destructive denoising filtering step by an additive generative step instead during the raytracing of the realtime rendering (plus doing things about subsurface and materials for now)
 
Last edited:
You must log in or register to reply here.
Back
Top