3 minutes ago, Republic3D said:

I have nothing against AI generated frames, if the feeling of fluidity and smoothness increases - I don't care if those frames come from my hard working CPU or my hallucinating GPU.

That’s where I’m at too. I guess if you’re a visual quality aficionado and go on a quest to actively look for distortions, DLSS3 will probably be a huge disappointment.

In terms of fps though, I’m happy if I can get around 60 fps most of the time. I’m in pancake mode with no head tracking, so around 60 is sufficient when flying fast and low or quickly panning around with the camera.

I found this video useful and interesting although it’s not specific to MSFS.

 

It's not machine learning, it's conditional vector math which comes from aggregated data which is just a different way of "interpolating" images, even though throughout Nvidia's papers you will hear them abusing the terms AI and adamantly claiming they are not using interpolation. Part of the algorithm has to use varying methods of interpolation, as comparing images is too slow even on modern hardware. Companies lie, marketing lies, etc... I'm not against any specific company (I still own Nvidia stock), but they have really jumped on the AI hype wagon in a bad way.

BTW, Nvidia was recently fined a few million for lying about how many cards they sold to crypto miners in their stock report, do you really think they are going to be honest about something that is almost impossible to disprove (whether or not they actually used machine learning to generate their final algorithm). It's possible (although highly unlikely) a few people at the company had some AI plugin in some predictive imaging application that helped them spot some issue with the algorithm, but there are almost ZERO useful machine learning algorithms when it comes to using ML to help optimize the way you optimize real-time image algorithms. The reason is because almost all methods of real-time image optimization are well documented, already known, and all in the public domain. Hence, that is why Nvidia basically used a tried and true old frame insertion method to gain extra frames. There are no scientists or research teams that successfully used predictive AI or learning AI to greatly enhance modern imagery in real-time, it's the reverse, it's all been experimental and lead to problematic outcomes. This does not mean AI hasn't been successfully used to advance imagery in longer running predictive fashion, but not in the way these companies are claiming it is. Hence, there are many fake papers on it too. One of the leading AI PHD researchers is now using most of his time to go around and re-explain AI to people and telling companies to stop lying about the way their algorithms are using AI. This is the most common and perpetuated lie in the imagery field, people have been using this lie for 20+ years to market their products. It's sad to see a big company attach themselves to this same lie, claiming all kinds of their imaging algorithms are "special" because they used Machine Learning to help generate them. This lie has been used by companies doing IPO's, many numerous scientists faking AI imagery in research papers (whom were later disgraced), and this stuff is well documented at MIT because it has happened so many times. This same ML type lie to improve imagery has been going on since 2002 or so, it's old and tiring. Because Nvidia is such a large company, they obviously do have their hands in experimental AI technologies, but it has very little do with their actual algorithms implemented on their cards.
 

Edited October 22, 20223 yr by Alpine Scenery

2 hours ago, Alpine Scenery said:

AI can be called anything, frame generation is not AI, sorry. This isn't AI as to how I would generally classify AI. 

You seem very knowledgeable about the term AI, so I’d like to ask your opinion on the fact that Digital Foundry refers to these GPU-generated frames as AI generated.

Quote

“This generated frame can be called an AI generated frame essentially, as a machine learning program on the GPU is deciding the information from these two images and how to combine them…”

Perhaps they’re just using the terms coined by nVidia themselves?

Edited October 22, 20223 yr by Cpt_Piett

I'm far from an expert at AI, I know some of the basic stuff, have coded with some of the plugins (just messed around with it). That said, I've read the imaging research on it by the most respected authors that are the "truth tellers" of what is and is not possible. We just aren't there yet with real-time image processing, it's just a shame NVIDIA jumped on the most generic AI lie that exists in the imagery field.

It's exactly the same type of processing all the AI researchers at MIT got in trouble for and eventually got kicked out of MIT. A lot of the fake papers were dealing with this exact topic, supersampling. One of the specific fake papers was based on upscaling, it claimed it could take a 480p image, supersample it up to 720p with no visual difference to a real image, then the claims got bigger and they took it all the way to 1080p. They created a fake paper claiming they could reproduce certain images in real-time, some guy even launched a software company based on the fake paper and embezzled a bunch of money from investors, but they were secretly using a database of "highly detailed" data that made the "fake AI" only work in certain situations, which is the opposite of what they claimed. I may be getting two stories mixed together, so some of my facts are combined and not accurate to one story (but across many). The old company Lucent got in big trouble for a similar story, using AI and paper thin like screens, he claimed they were going to bring TV's to all our cereal boxes for advertising revenue. They also made some adamant claim about he had some supersampling like technique to make a processor that could be simpler than a $5 calculator uses to render graphics in real-time. As you can imagine, it artificially knocked the stock up by several % points before the company went bankrupt and the CEO ended up in prison. His lie was actually based on the "experimental at the time" idea of "rollup" oLED screens I believe, which may one day come to consumers at dirt cheap prices, in like 10 to 20 years from now (or 30+ years from when he made the claim). His claim was Lucent had the technology to make it possible in 5-10 years, but that is not why he went to prison (it was for something else). That had to do with some type of plastics company scam deal, but it just goes to show you these companies aren't always telling the truth (and that's an understatement).

Not to say supersampling and frame interpolation isn't a good technique, but because it's a good technique, is why the companies latch onto it and exaggerate its capabilities and make claims like it uses Machine Learning.

As noted earlier, Sony has been claiming they were using Machine Learning with Reality Creation I think it was called, it was a total lie. There was ZERO machine learning in the algorithm, as it was just a basic sharpening algorithm. Unless a company is specifically hoodwinking investors based on one claim and has no revenue, the regulators don't care what they say. Only small companies getting new investors have to watch what they say (mostly), big companies just get fined. The regulators aren't going to dismantle a large company based on a few marketing fibs.

Edited October 23, 20223 yr by Alpine Scenery

Everyone knows what a butterfly is. I see a butterfly I know its a butterfly. Festo makes artificial butterfly's, I guess for promotional reasons. When I see the Festo mechanical butterfly. I recognize it as a butterfly but I know its artificial. It's reasonable to call it an artificial butterfly. An AB if you will. Even though it only moves like and resembles a real butterfly. It has nothing else in common with a real butterfly. Intelligence is a far more ethereal thing than a butterfly. Even among philosophers there is a great deal of disagreement as to what Intelligence is. It seem to me that Festo had to have a really good idea of what a butterfly is before they could do such a good job of making a fake one. It would fool anyone from a sufficient distance if they didn't know it was fake. The test for artificial intelligence is the Turing test. For an AB to pass the Turing test it would have to fool a real butterfly.  I don't think any machine has ever passed the Turing test therefore there is no such thing as artificial intelligence that we know off. A.I. and Unicorns have the fact that they don't exist in common.😁

10 minutes ago, FBW737 said:

It would fool anyone from a sufficient distance if they didn't know it was fake. The test for artificial intelligence is the Turing test. For an AB to pass the Turing test it would have to fool a real butterfly.  I don't think any machine has ever passed the Turing test therefore there is no such thing as artificial intelligence that we know off. A.I. and Unicorns have the fact that they don't exist in common.😁

You make a good point, but using the specific terminology "Machine Learning" implies a specific type of algorithmic process, which is definitely not likely to be used in a real-time super-sampling technique. Again, Sony has been making this claim since like 2005, if your familiar with AVSForum, someone in their had written a history of Sony's marketing AI, and it was pretty funny. NVIDIA is more convincing because they have so much power in the card that it almost sounds true, but in reality, the only AI in any of their algorithms are mostly just conditional type stuff (and that stuff can be debated to whether or not it's AI, but it certainly isn't ML).

If someone wants to call something AI that does advanced image recognition or something, fine I can live with that. BUT... When people latch onto the specific terminology of machine learning (as Sony and others did in the past), that is going too far and is an outright lie. In reality, these imaging algorithms are 20+ years old just running on newer hardware, and many of them are the farthest thing in the world from AI, hence they are the exact opposite. Rather than them being computationally non-determinate until the final decision cycle, they are completely pre-determined algorithms that are just plug and play math equations (And this is the LIE of it)...

As far as people not being able to determine what AI is, some judges would disagree, since the guy ended up in prison for specifically hoodwinking people about what AI could do with imagery.

 

Edited October 23, 20223 yr by Alpine Scenery

1 hour ago, Alpine Scenery said:

It's not machine learning, it's conditional vector math which comes from aggregated data which is just a different way of "interpolating" images, even though throughout Nvidia's papers you will hear them abusing the terms AI and adamantly claiming they are not using interpolation. Part of the algorithm has to use varying methods of interpolation, as comparing images is too slow even on modern hardware. Companies lie, marketing lies, etc... I'm not against any specific company (I still own Nvidia stock), but they have really jumped on the AI hype wagon in a bad way.

BTW, Nvidia was recently fined a few million for lying about how many cards they sold to crypto miners in their stock report, do you really think they are going to be honest about something that is almost impossible to disprove (whether or not they actually used machine learning to generate their final algorithm). It's possible (although highly unlikely) a few people at the company had some AI plugin in some predictive imaging application that helped them spot some issue with the algorithm, but there are almost ZERO useful machine learning algorithms when it comes to using ML to help optimize the way you optimize real-time image algorithms. The reason is because almost all methods of real-time image optimization are well documented, already known, and all in the public domain. Hence, that is why Nvidia basically used a tried and true old frame insertion method to gain extra frames. There are no scientists or research teams that successfully used predictive AI or learning AI to greatly enhance modern imagery in real-time, it's the reverse, it's all been experimental and lead to problematic outcomes. This does not mean AI hasn't been successfully used to advance imagery in longer running predictive fashion, but not in the way these companies are claiming it is. Hence, there are many fake papers on it too. One of the leading AI PHD researchers is now using most of his time to go around and re-explain AI to people and telling companies to stop lying about the way their algorithms are using AI. This is the most common and perpetuated lie in the imagery field, people have been using this lie for 20+ years to market their products. It's sad to see a big company attach themselves to this same lie, claiming all kinds of their imaging algorithms are "special" because they used Machine Learning to help generate them. This lie has been used by companies doing IPO's, many numerous scientists faking AI imagery in research papers (whom were later disgraced), and this stuff is well documented at MIT because it has happened so many times. This same ML type lie to improve imagery has been going on since 2002 or so, it's old and tiring. Because Nvidia is such a large company, they obviously do have their hands in experimental AI technologies, but it has very little do with their actual algorithms implemented on their cards.
 

Thank you for an interesting post. It's good to hear multiple opinions on matters like these. I don't know much about AI except for the AlphaGo thing, and using it in image and video editing both via apps on my phone but also in Photoshop Beta. It's advancing rapidly. Oh and in DALL-E and Midjourney. That's some seriously impressive stuff. 
I have high hopes AMD will be able to match or surpass Nvidia with regards to this tech.

14 minutes ago, Alpine Scenery said:

As far as people not being able to determine what AI is

It’s an interesting discussion. Somewhat tangentially related (perhaps to AI but we’re getting further away from the OP topic), is the intriguing story this summer about the Google technician who believes that LaMDA had become sentient. Google has refuted this by claiming its impossible, and apparently they’ve hard-coded it to prevent it from passing the Turing test. 

AI is an amazing thing, but it has limitations in real-time applications, imagine expecting Photoshop to complete every task you ask it to do within .017 seconds, which is approximately the amount of time you have in real-time processing to finish preparing a frame before you induce lag at 60fps, and this is BEFORE taking into account your pipeline lag. Latency has been the Achilles hill of real-time image processing forever, and will be for a good bit while longer.

 

Myself I did not dig into it (yet) so in simple words if I am correct : first downscaling, then upscaling and adding an AI frame between 2 real frames so the framerate doubles …

What is important for me : at a large airport with many AI ac and bad weather where my 12900K cpu is the bottleneck and goes down to 37/38 fps , will DLSS 3 bring extra smoothness because of the extra AI frames ?

We all know the jumping from the scene when having just 10 fps or even with 30 fps and using a fast ac, making a fast turn left and look left out the window.

Just curious as I might pick up a Inno3D X3 4090 + Dellta GPS 1300W Ps today …

 

Edited October 23, 20223 yr by GSalden

It's rendering at the lower resolution from the get go just as if you had set the game to run in a lower resolution. Then it upscales it with supersampling and uses a frame insertion technique (which is a form of interpolation, because using pure aggregation to insert a frame doesn't make any sense), companies just flat out lie about this stuff, sad to say but true. As one final example, Sony actually had convinced people that their reality creation engine used AI on the textures themselves and referenced a database using AI to texture match to improve the image in movies (that would be like tech you might see in 2060), but it was all entirely made up fantasy by their marketing dept. You could potentially achieve close to the same relative effect by just rendering in a lower resolution, and compensating with other settings, minus maybe some of the smoothing from the FI, and maybe not get quite as good anti-aliasing (which is the main benefit of super sampling).

Smoothness can be perfectly faked in the resulting images only if the imagery is slowed down in a smooth enough manner, hence equally slowed without too much judder. Frame Interpolation can be better or worse depending upon specific corners cut or specific implementations of the algorithm, so I'll at least give Nvidia the benefit of the doubt here and say since they have extra processing power available, maybe theirs is slightly better than most. Honestly, it doesn't sound worth the upgrade on this feature alone, the feature has been oversold IMO, but make your own decision.

Edited October 23, 20223 yr by Alpine Scenery

Then I came across this vid :

 

Some of the things :

- DSLL3 does not work on reflections and moving shadows 

- DLSS3 causes lag which shows especially in compatitive  games 

 

Edited October 23, 20223 yr by GSalden

I also saw that video above you posted, and the conflicting reports had pretty much confirmed to me that their was an inner battle within Nvidia between the marketing stretching the truth, and the engineers trying to talk their way around the mess that the marketing had created. They got caught up in a bunch of contradictory statements. AFIK, it's literally impossible to generate frames with AI in something as "real-time" as a GPU, every single millisecond counts because you are trying to maximize FPS. 

This guy does a pretty good analysis starting at 11:15...
 

 

9 minutes ago, Alpine Scenery said:

I also saw that video above you posted, and the conflicting reports had pretty much confirmed to me that their was an inner battle within Nvidia between the marketing stretching the truth, and the engineers trying to talk their way around the mess that the marketing had created. They got caught up in a bunch of contradictory statements. AFIK, it's literally impossible to generate frames with AI in something as "real-time" as a GPU, every single millisecond counts because you are trying to maximize FPS. 

This guy does a pretty good analysis starting at 11:15...
 

 

Thanks for the vid. You can see the flaws when panning. 
 

In my case I use a fixed view but when making turns the view looks like panning of coarse.

I am curious what current owners experiences are. We have a few with a 4090 ….

1 hour ago, Alpine Scenery said:

I also saw that video above you posted, and the conflicting reports had pretty much confirmed to me that their was an inner battle within Nvidia between the marketing stretching the truth, and the engineers trying to talk their way around the mess that the marketing had created. They got caught up in a bunch of contradictory statements. AFIK, it's literally impossible to generate frames with AI in something as "real-time" as a GPU, every single millisecond counts because you are trying to maximize FPS. 

This guy does a pretty good analysis starting at 11:15...
 

 

Who cares real or not real!!!!

What matters is that games play superb. 

Also very happy with frame generation on in combination with FS2020.  
Also F1 feels like the real deal with frame generation on.  
 

I don't care what all the vlogs say!!!
I care about my own experience!!

Same with food vlogs telling us a Big Mac is not very tasteful.  

Well I will decide about that. 

I also think that if the 4090 was 900 euro everyone would be happy and where not bashing against the 4090. 

 

 

Edited October 23, 20223 yr by altenae