Monday, January 6th 2025

NVIDIA Introduces DLSS 4 with Multi-Frame Generation for up to 8X Framerate Uplifts

With the GeForce RTX 50-series "Blackwell" generation, NVIDIA is introducing the new DLSS 4 technology. The most groundbreaking feature being introduced with DLSS 4 is multi-frame generation. The technology relies on generative AI to predict up to three frames ahead of a conventionally rendered frame, which in and of itself could be a result of super resolution. Since DLSS SR can effectively upscale 1 pixel into 4 (i.e. turn a 1080p render into 4K output), and DLSS 4 generates the following three frames, DLSS 4 effectively has a pixel generation factor of 1:15 (15 in every 16 pixels are generated outside the rendering pipeline). When it launches alongside the GeForce RTX 50-series later this month, over 75 game titles will be ready for DLSS 4. Multi-frame generation is a feature exclusive to "Blackwell."
Add your own comment

135 Comments on NVIDIA Introduces DLSS 4 with Multi-Frame Generation for up to 8X Framerate Uplifts

#126
londiste
Capitan HarlockDepending on the game 4K does smooth edges on is own without any AA needed so that's already not true that is necessary always.
No, resolution has absolutely nothing to do with edge aliasing.
Like I said before, your perception of aliasing might, but that is a different topic. And not a thing with normal monitor pixel densities.
Capitan HarlockRtx is not necessary always because is just like Physx used to be a thing in the past.
Physx is a whole different topic. And it did not go away because the idea was bad but primarily because CPUs got more and more cores and physics engines that could be well enough parallelized got much more performance to play with. Nvidia hobbled Physx for a while so that the version of it running on CPUs was single threaded but in the end gave up and made it work properly. Physx itself did not go away, the forced-to-run-on-GPU variant did.
developer.nvidia.com/physx-sdk
en.wikipedia.org/wiki/PhysX#PhysX_in_video_games
Posted on Reply
#127
Blueberries
SOAREVERSORThe 2000 ask is not for gamers. It's for pros. The 1000 ask is for the inbetween. The below are for gamers. People need to get that through their skulls and fast. There is a reason AMD is not competing with those because those are not gaming cards. The companies are telling us this. Yet PC gamers refuse to accept what the specs, the price, and the companies themselves keep saying. Because they are spoiled fucking shits.
That's just objectively false. Anything "RTX" is a gaming card and will not work with professional drivers.
Posted on Reply
#130
SOAREVERSOR
BlueberriesMy mistake then, that must have changed in recent years.

That's not to say the RTX 5090 is not marketed as a gaming card though.
It's marketed as a prosumer card. CUDA does not need drivers. Quadro drivers are only for CAD CAM. You're either a liar or a fucking idiot. 5090 is a prosumer card. Hell even 5080 straddles that line. Bellow that are the gaming cards. Get a clue.

Hint CUDA is a toolkit. Usually ran on linux. developer.nvidia.com/cuda-downloads that right there is the justification for the 5090. You might have missed it. That was the justification for the 8800 GTX. It's only got more extreme since Titan. That's why GPUs get banned to China. Not because 5090s will let their gold farmers in WoW out gold farm our gold farmers.

Rasterization is dead. You're buying an AI product and it's fakery. Make no mistake about it.
Posted on Reply
#131
Whitestar
SOAREVERSORThe 2000 ask is not for gamers. It's for pros.
I present to you the Nvidia 5090 page, tadaaaa! GeForce RTX 5090 Graphics Cards | NVIDIA
It's the fastest gaming card ever. It's marketed towards gamers. Please explain how it's not a gaming card other than "The companies are telling us this".
SOAREVERSORIt's marketed as a prosumer card. You're either a liar or a fucking idiot.
Again, in what way would you say they are not trying to attract gamers with that page? Good luck!

Oh and that last part; I have a feeling you'd better apologize or delete it.
Posted on Reply
#132
igormp
BagerklestyneNothing looks better than native, can look as close to native as possible, can't look better. Can't. It's interpreted math. That's not progress (getting closer is.) That's like saying AI upscaling a picture makes it look better at the same resolution.
I agree with your word, but disagree with what you meant. Reminder that the only actual "native" thing for us would be you actually seeing something with your own eyes.
Renders are just some math that try to approximate what we have in the real world. Our current approximations for games (in raw raster) is reasonable, but far from good given the real time constraints. We can achieve better quality with longer renders that do more math, and more precise math (that are still far from real life, but improving), but that'd be useless for games given the speed.

Upscaling with machine learning is nothing more than another different way to try to do this approximation in a non-deterministic way.
BagerklestyneThing is, that's taken the sub optimal time degraded image and restored it.

If you took a picture of the same subject today with a quality modern camera at the same detail level (pretend for arguments sake the original was 6000x4000 pixels - so a camera that uses the same resolution) then you couldn't improve upon the picture taken without interpolation, which is not the original image then.
Any representation that's not the real thing is a sub-optimal representation still. Even a modern camera at 6000x4000, even though it's considered good nowadays, still can't capture all details present.
That same 6000x4000 in some years will also look like a graded image that we will find ways to improve with future tech.
BlueberriesThat's just objectively false. Anything "RTX" is a gaming card and will not work with professional drivers.
Their professional lineup is also called "RTX", without the geforce branding since they dropped the "Quadro/Tesla" monikers:
www.nvidia.com/en-us/design-visualization/rtx-6000/
SOAREVERSORIt's marketed as a prosumer card. CUDA does not need drivers. Quadro drivers are only for CAD CAM. You're either a liar or a fucking idiot. 5090 is a prosumer card. Hell even 5080 straddles that line. Bellow that are the gaming cards. Get a clue.

Hint CUDA is a toolkit. Usually ran on linux. developer.nvidia.com/cuda-downloads that right there is the justification for the 5090. You might have missed it. That was the justification for the 8800 GTX. It's only got more extreme since Titan. That's why GPUs get banned to China. Not because 5090s will let their gold farmers in WoW out gold farm our gold farmers.

Rasterization is dead. You're buying an AI product and it's fakery. Make no mistake about it.
FWIW CUDA does need drivers. And on linux it's a single driver for all product lines, be it quadro, tesla or geforce (only the tegra ones get a different driver).
Posted on Reply
#134
Dawora
N3utroI'd bet these people saying they dont like "fake frames" wouldn't be able to tell which is the AI generated frames gameplay vs non AI generated at the same framerate in a blind test. Probably would say the AI generated gameplay would even look better than the original one at the same framerate.
ppls say because they use AMD and Nvidia is now Rank#1 in FG/Upscaling
So they dont like FG or DLSS because AMD clearly is 2# now

If Amd is better than Nvidia in FG/upscaling Amd fans will love it.
Posted on Reply
#135
Krit
A little bit closer to the truth!

[COLOR=rgb(226, 80, 65)]I was right about the 5090. It’s BAD[/COLOR]

Posted on Reply
Add your own comment
Jan 9th, 2025 10:58 EST change timezone

New Forum Posts

Popular Reviews

Controversial News Posts