Talk:Deep learning super sampling

Article is self-contradictory, lacks verifiability, uses vague, non-descriptive language and is written like an advertisement
The article:


 * Contradicts itself by both claiming that the technology is based on deep learning while simultaneously claiming a game with the technology was released without deep learning. Appears to show a deep issue with the sources the article uses or the subject itself.
 * Is not verifiable. Core claim of the article is not in the citation given (the claim that by using this technology quality is same with lower resolution than a higher resolution without). Thus the article fails WP:VERIFY. Further it should be questioned if this claim is a fact or an opinion in the first place.
 * Uses vague and non-descriptive language. "Video card overhead" is not a meaningful definition of something and does not yield meaningful definitions by searching.
 * Uses weasel words, such as in "is said to use machine learning" WP:Weasel -Weasel words are words and phrases aimed at creating an impression that something specific and meaningful has been said, when in fact only a vague or ambiguous claim has been communicated
 * Fails to describe the technology from neutral point of view by just presenting it as lower overhead and same quality with lower resolution. Rather it qualifies as advocacy, and as such the article is written like an advertisement as described in WP:PROMO.

I'd suggest this for deletion. 62.248.185.87 (talk)
 * Why wanting to remove this article? Try to improve it instead. I tried to add a lot of reliable sources. Hervegirod (talk) 09:10, 14 April 2020 (UTC)
 * First remark: It may be possible to better clarify, but the first DLSS 1.0 had deep learning according to the sources, the second one (not named) had not deep learning anymore, and the current one has deep learning. I think I clearly stated it in the article. For example: "which this time is said to use machine learning...." Hervegirod (talk) 09:10, 14 April 2020 (UTC)
 * Second remark: Maybe trying to get a better wording? What do you propose? Hervegirod (talk) 09:10, 14 April 2020 (UTC)
 * Last remark: advocacy? I tried to get as many reliable sources as I could, the sources all criticized the first versions of DLSS saying that it was not working which is what I wrote in the article, but they for the moment all said that the last version is working very well. Hervegirod (talk) 09:10, 14 April 2020 (UTC)
 * Appears to me that this article is based on NVidia's claims, and buzzwordy jargon. It is so hip now to do machine learning or deep learning. Does this merit an article separate from Supersampling? Pikavoom (talk) 09:25, 14 April 2020 (UTC)
 * I don't think so, the technology appear to be very different from "simple" Supersampling, the method it uses is completely different. Including this article in the other would make the first one very difficult to understand. And I did not use only Nvidia sources, but a lot of Kotaku, techquila, Forbes, hothardware, Eurogamer, Anandtech sources. Do you think these sources are not reliable? Again if you want to improve the wording, please do so. Hervegirod (talk) 09:33, 14 April 2020 (UTC)
 * "In 2019, the videogame Control shipped with Ray tracing and an improved version of DLSS, but which didn't use Deep learning": you have put "contradictory", but it's exactly what the source say, the exact text is: "Of course, this isn't the first DLSS implementation we've seen in Control. The game shipped with a decent enough rendition of the technology that didn't actually use the machine learning"(which I quoted). It's exactly what I wrote in the history section. Hervegirod (talk) 09:39, 14 April 2020 (UTC)
 * NVidia's claims are reliably sourced, but these sources are mostly rehashing NVidia PR. As for the method, how is this a non-trivial combination of Deep learning and Supersampling?Pikavoom (talk) 09:43, 14 April 2020 (UTC)
 * For example the hothardware source, or Anandtech, did not just rewrote what Nvidia said, but made their own benchmarks. And about the fact that it would be a trivial combination of Deep learning and Supersampling, two thirds of the Supersampling article in the supersampling article talk about the method or algorithm, the method in DLSS has nothing in common with the one described in the Supersampling article. It does not work by using AI to execute the Supersampling algorithm at all. Hervegirod (talk) 09:49, 14 April 2020 (UTC)
 * Yes, but that is a benchmark of GeForce RTX 20's implementation of DLSS, not the really the technique itself.Pikavoom (talk) 09:52, 14 April 2020 (UTC)
 * DLSS is only available on this family of video cards, it's for the moment specific to these Nvidia cards. Hervegirod (talk) 12:42, 14 April 2020 (UTC)
 * "Benchmarks on Control tend to show that the resulting image at a 1080 pixels resolution upscaled from a 720 pixels resolution have the same quality as a native 1080 pixels resolution but retain the 720 pixels resolution performance" I think I forgot to put the source I found for this one, I will search for it. Hervegirod (talk) 09:45, 14 April 2020 (UTC)
 * I don't remember where I found it, I replaced by an equivalent PC gamer one, which is also based on a very recent benchmark performed by PC gamer.Hervegirod (talk) 09:54, 16 April 2020 (UTC)
 * If after about a month and half you don't remember where you've found a key argument for a technology for which you've exclusively written an article for, and the citation you've given for it directly contradicts the claim you've put in, then the claim clearly can't be in the article. 62.248.185.87 (talk) 17:29, 26 May 2020 (UTC)
 * Do you ever read what people answer to your questions? I explained that I replaced the part where I forgot the source by a sourced PC gamer article. It appears that you now removed it. Your behavior is bordering vandalism. Hervegirod (talk) 23:45, 26 May 2020 (UTC)
 * Neither of the claims you wrote in that sentence was in the cited article, in fact the article pretty much said neither of them were true. You wrote that the performance doubled on the game control on quality preset double the native resolution performance, the article showed it did nothing of the sort - on the table in the article it says performance moved from 57 fps to 91. You wrote that the quality was the same resulting image... have the same quality as the native - the article went out to point some of the shortcomings, such as Hair's still a sticking point, however, and aliasing where the AI has struggled to upscale the source material and while the article does go on to conclude there were little to no expense, particularly the little part of that is not the same as the statement same quality, which you insist on putting on the article. Furthermore none of the testing in the supposedly source material of these claims even is at 4k output resolution! Rather description under the table says Control benchmarks were carried out at 1440p. What is the actual source for your claims? It clearly isn't this article. Also I already noted most of this in the edit summary, which you've conveniently ignored only to go on to rather launch further attacks against my character, which at this point include asking "do you ever read", accusations of vandalism, and stating that I had not read an article which I've spent hours commenting on. These repeated baseless accusations while refusing to read or address edit summaries is not even remotely close to good faith participation. 62.248.185.87 (talk) 09:29, 27 May 2020 (UTC)

Regarding the suggestion this article should be deleted: It's reasonable to raise that issue here but when there is opposition the only recourse would be to nominate the article for deletion at WP:AFD. Such a nomination would almost certainly fail because Google shows lots of different sources talking about the term in the context used by this article. Accordingly, notability is established and talk about deletion is a waste of time. As an uninvolved administrator, I am merely reporting standard procedure and have no opinion on the merits of this article. Johnuniq (talk) 07:05, 30 May 2020 (UTC)
 * There is still a [contradictory] tag in the article. I am trying to focus attention on the issues while staying uninvolved so am asking if contradictory still applies. The justification for the tag appears to be "Contradicts itself by both claiming that the technology is based on deep learning while simultaneously claiming a game with the technology was released without deep learning." The article says "Nvidia advertised DLSS" at launch in 2018 and a particular game was issued in 2019 where deep learning was not used. Is the tag justified? Johnuniq (talk) 07:16, 30 May 2020 (UTC)
 * Hello, I tried to discuss the remarks about deep learning in this talk page Hervegirod (talk) 13:13, 30 May 2020 (UTC)
 * Yes, the article is still contradictory. The very first sentence of the article states that the technology is deep learning technology "(DLSS) is a technology developed by Nvidia, using deep learning", where as the release history states that the "2.0 (first iteration)" does "not use machine learning", which [[Media:AI-ML-DL.png|by definition includes deep learning]]. Also I would not propose deletion anymore. 62.248.185.87 (talk) 19:30, 30 May 2020 (UTC)
 * As of now / after Dicklyons revision 959815010 added "(in some versions)" to the article, it doesn't contradict itself, but I personally don't feel confident in removing the tags since I believe it will take a miracle for the remark to stay in the article for as long as until it can be clarified with a reliable and sourced reference and not just edit warred out of the article. I base this on the experience of how removing a single sentence entirely contradicted by it's citation out of this article got me character assassinated on my user talk and a mod talk page, how my revisions concerning 2716 word referenced article were undone seconds after making them while none of the concerns were or ever have been addressed, and worst of all these behaviors only lead to actions against me. 62.248.185.87 (talk) 21:07, 30 May 2020 (UTC)

103.27.230.134 (talk) 04:15, 19 August 2020 (UTC) DLSS 2.0 is still based on AI. What made the editor think DLSS 2.0 for Control doesn't use Deep Learning. It just says that DLSS 2.0 doesn't need specific training for each game. 103.27.230.134 (talk) 04:15, 19 August 2020 (UTC)

Contradictions fixes
Previously one editor said that the article contradicted itself because DLSS is said to be based on AI, but the second version was not based on AI. It was not clear enough, partly because they did not change the release version between this one and the last recent one (the two of them are all named DLSS 2.0 by Nvidia). It think I clarified it, also because I added a history chapter. So I removed the banner about the clarification at the top of the article. Hervegirod (talk) 09:57, 16 April 2020 (UTC)


 * Why are you removing contradictory tags when the article still contradicts itself? The article can't at the same time claim in the lead section that it's a technology based on Deep Learning, [[Media:AI-ML-DL.png|which is a form of AI]], while insisting later that a game shipped with the technology which doesn't use Deep learning or AI for it. I've added the contradictory remarks back in. 62.248.185.87 (talk) 10:18, 26 May 2020 (UTC)


 * Maybe you should read the article and the sources, it is explained, but you obviously did not. It is not contradictory. Hervegirod (talk) 15:15, 26 May 2020 (UTC)


 * I read the article. The article can't at the same time claim the technology is based on deep learning and then say that it's not. Only one of these statements can be true at the same time. Saying it's technology based on deep learning is exclusionary language that closes the possibility to say later that it's not based on deep learning. Also your method of participating in the conversation is not in good faith - there was no reason to accuse me of not having read an article I've commented this much on. 62.248.185.87 (talk) 16:32, 26 May 2020 (UTC)
 * See Hervegirod (talk) 13:06, 30 May 2020 (UTC)

Weasel words
The sentence is: "which this time is said to use machine learning and don't need to be trained on every game it is applied to". If somebody could find a better wording... Hervegirod (talk) 09:58, 16 April 2020 (UTC)
 * I changed the wording. Hervegirod (talk) 14:53, 16 April 2020 (UTC)
 * I have put again the weasel words tags in front of the text because I mistakingly removed it while trying to improve the text. I don't think that it still valid, but then people can assess with the current sentence. Hervegirod (talk) 07:46, 17 April 2020 (UTC)
 * I now changed the wording on this sentence, so ultimately removed the tag. Hervegirod (talk) 08:45, 19 April 2020 (UTC)

About Nvidia claims
DLSS 2.0 seems to be working really well with the last 2.0 version, as it has been confirmed by independent and reliable sources which did their own benchmarks. However, what we still don't know is if the technology is really working well on any game which use the Nvidia provided API, or if it only works on the specific games which enable it for now. Nvidia explained that DLSS 2.0 is not trained specifically on every game, but as only a few games use the latest version of this technology, we still don't know if it works effectively on any game which use it. I would have added this on the article, but I can't for the moment, because I could not find any source for that. Only time will tell if the "generic" term used by Nvidia is really true. Hervegirod (talk) 10:42, 19 April 2020 (UTC)

The phrase "video card overhead", again
The phrase "video card overhead" has no established definition, and does not yield meaningful search results.

Searching for the phrase in google yields me 5 pages of results and total of 48 links.
 * The first link gets you straight back to this article
 * 26 are copies from this article like Wikipedia mirrors or short copies for SEO spam purposes
 * 5 do not use the words together but google has picked due to there being a line change or otherwise (such as in a formatted document)
 * 3 are pure SEO spam without being copies
 * All in all 73% of the results are the article or copies of it, SEO spam, or mistaken results
 * 2 of them is the same comment to physical clearance from a connector
 * 1 refers to video card memory usage in a unclear manner, either way an unrelated concept
 * 1 refers to saved resource usage in a headless installation
 * 1 refers to non-utilized performance of a video card while the video card still is performing something
 * 1 refers to delay in a video stream caused by a video card
 * 1 refers to the general concept of running software consuming performance
 * 1 is about someones computer crashing if their video card is stressed

There's one single reference in a book about 3d modelling software Maya seemingly using it in the way used in the article, which appears 4 times. But even that doesn't really specify what it means more than as lost performance.

Curiously there is a single reference in a gaming access weekly article written 3 days ago about this technology, however this article doesn't define it.

This seems to count as WP:Weasel as well, but whether it does or doesn't either way I don't think matters as it's not language that should be used on a Wikipedia article as it doesn't convey meaning. 62.248.185.87 (talk) 22:18, 29 May 2020 (UTC)


 * I replaced "overhead" with something that might work better. Dicklyon (talk) 05:23, 30 May 2020 (UTC)
 * Thanks, although as you would know, the original is standard terminology: Overhead (computing). Johnuniq (talk) 06:59, 30 May 2020 (UTC)
 * Yes, I'm aware. It doesn't seem to be the right term/concept here, though I note that a source did use it. Dicklyon (talk) 15:47, 30 May 2020 (UTC)
 * It's not really the same as the articles examples. For example, based on this article, "protocol overhead" is the overhead taken by a protocols control and signalling data. So if data is sent in a bitstream with no protocol control and signalling data then there's no protocol overhead. The phrase is almost self explanatory, all one needs to know is what protocol data is. In stark contrast knowing what a video card is doesn't help you understand what the phrase "video card overhead" means, and removing the video card certainly wouldn't be the way to get rid of whatever it is that is referred to by the phrase here. It certainly isn't self-explanatory and doesn't seem to make much sense. 62.248.185.87 (talk) 18:55, 30 May 2020 (UTC)
 * Overhead is talking about the amount of work the video card has to do, and that greatly affects the speed at which graphics are drawn and the power used (heat that needs to be dissipated). Johnuniq (talk) 03:54, 31 May 2020 (UTC)

Remark about Deep learning / Not Deep learning issue
I just read the contradict tag added again by User talk:62.248.185.87 with this summary : "If they called it DLSS before it had deep learning then the article can't say it's a technology based on deep learning". It's not me who called it "Deep learning super sampling" but Nvidia, furthermore the sources I found state that: I tried to clearly say and source that in the article, especially on the "Release history" chapter. It's not my fault if Nvidia was not always clear on their promotion of their own technology. Maybe the wording could be improved to make it clearer (it seems clear in this article for me, but I am the one who wrote the "Release history" chapter originally) Hervegirod (talk)
 * It used deep learning for DLSS 1.0, and learning was processed on a per game basis. So when Nvidia first promoted their technology, it used deep learning
 * The first version did not work well, so Nvidia did not used Deep learning anymore for the first iteration of DLSS 2.0. I'm writing first iteration  because it seems that they did not change the release number of their technology between this one and the current one. However the sources explain why this version was very different from the "current" DLSS 2.0 one
 * The current version uses deep learning again, this time with a generic learning not specific to a game, for the current iteration of DLSS 2.0


 * In summary what you have described is that DLSS is not necessarily based on deep learning. Thus describing it as a technology which uses deep learning, as is done in the lead section, is not correct. Nvidias failure to name only technologies that use deep learning as DLSS is perhaps a problem for their marketing team, but not for Wikipedia, and means this article can't say DLSS is based on deep learning in the introduction.


 * Also I don't think anyone claimed it's your fault. My edit summary was in response to earlier edit by Dicklyon, which removed the tag with the commentary "yes they might have called it DLSS even before it had DL" - hence the response "If they called it DLSS before it had deep learning then the article can't say it's a technology based on deep learning".


 * If Wikipedia allows technologies which are not necessarily deep learning based to introduce them as deep learning then soon half of the technology articles will be saying they are deep learning technologies. 62.248.185.87 (talk) 16:57, 30 May 2020 (UTC)


 * That's not up to us to decide, many sources say that the first and last iteration were / are deep learning based. Hervegirod (talk) 17:31, 30 May 2020 (UTC)


 * Whether the last and first iteration were deep learning based was never under contention. You yourself added the remark that a version was not.


 * Contradictory should remain as long as the article contradicts itself. Currently it does so by introducing DLSS as a deep learning technology and then later says there's a version which isn't. 62.248.185.87 (talk) 17:44, 30 May 2020 (UTC)

DLSS 2.1
Information about DLSS 2.1 should be added. It apparently adds an ultra performance mode as a major differentiating factor from previous versions. Svetroid (talk) 20:31, 8 December 2020 (UTC)

These two paragraphs... smh
I'll start before somebody goes fanboy on me by saying that I could care less which of the two video card companies is "winning" the game of screwing people into buying new cards / switching brands, but I have to comment on this part:

Tensor Cores are available since the Nvidia Volta GPU microarchitecture, which was first used on the Tesla V100 line of products. Their specificity is that each Tensor Core operates on 16 bits floating point 4 x 4 matrices, and seem to be designed to be used at the CUDA C++ level, even at the compiler level. Aside from the bizarre grammar, "operates on..." is a given, the real specificity (and how there are so many of them) is that they can do approximately jack squat besides multiply, add, load, and store elements to produce said matrices. This has no immediate appeal to the consumer market.

Fabrication of a chip for something like the V100 produces lots of defectives which they needed something to do with since enterprise customers won't accept the board where 1/4 of the SMs won't produce accurate results and another 1/4 won't work at all when they're doing something critical like drug discovery or fusion plasma simulation,  hence the invention of features like this where the LSB being wrong in an element 1/10000 times won't be noticed. Slap some RGB lights on it and you can market your e-waste to gamers. AMD and Intel do the same thing, so I'm not picking on anybody here.

"Designed to be used at the CUDA C++ level" is kind of a given since nVidia wants people locked in to CUDA, which again, everybody else tries to do, although AMD at least stuck with open standards.

"or even the compiler level" means absolutely nothing. You don't "use" the machine languages you're compiling for at the compiler level. You compile high level languages through possible multiple stages until you're at either target-specific assembly or binary machine code. Every processor feature in history (since HLLs have existed) was designed to be used at the compiler level, because writing programs in hexadecimal is painful even if you know the instruction set encoding like the back of your hand. Most people don't know or bother with learning assembly, either, so a feature that can't be implemented as a compiler feature / optimization might as well not exist. This is why Itanium was a failure, but I've already ranted enough and won't go into that.

Features like the tensor units that would require an expert at multithreaded programming to program manually and require constantly updated system state to run optimally aren't even left to the compiler, except to simplify things a little for the end user, who will then use a python module that interfaces with the c++ cuda libs because all those semicolons in c++ are too confusing for them. Then they'll release a pixel art metroidvania that requires tenorflow that only runs on one processor core and can't maintain 30fps at 1080p and complain about computers being too slow. See below for how the instructions actually work.

The Tensor Cores use CUDA Warp-Level Primitives on 32 parallel threads to take advantage of their parallel architecture. A Warp is a set of 32 threads which are configured to execute the same instruction.

Sorta...  it's an insanely verbose instruction (but a warp primitive in cuda compiles down to multiple instructions as below) that's sent to the warp scheduler to be executed as a ton of individual instructions on 32 hopefully optimally scheduled and localized (no guarantees are made) tensor subunits, then reassembled into a result when they finish. This cuts down on silicon per multiply-add subunit of the tensor cores (no need to do anything but multiply and accumulate, and a couple of simple bitwise ops on integers) but as far as i can tell you're SoL if you want a 4x4 matrix multiply add, the smallest "shape" they list is 8x4. There is no way, afaik, to individually address these sub-processing units from anywhere but the warp scheduler on chip, except maybe overriding data distribution. This is what the assembly looks like for a 16x16 matrix multiplication, which requires 32 tensor units each scheduled internally to do the required series of 4x4 operations:

So that's the level it's *designed* to be used at. Since most humans can't smoke enough meth without dying to want to write it like that, compiler support for the PTX instructions is constantly worked on and a CUDA API created. I went through that pedantic mess because the wording implies that being designed for use in c++ (it was either that or C) and goes on with "wow you can even use it in compilers as if that's some super elite feature or even makes any sense.  The miracle would be if it's supported by any compilers that nVidia employees didn't add the feature to.   As a former compiler engineer I'd rather mate with a garbage disposal than have to implement just the backend encoding for their instruction set, let alone an assembler / assembly printer or any kind of language support.

The big snafu as far as i can tell is that GPUs don't do conditional speculative execution in the normal sense. If there are multiple code paths, they execute all of them at once and throw away the results for anything not needed. This makes the AI / deep learning sound fancy but it's really a way to minimize the number of times that happens on a user machine by pre-deciding most things and letting the computer run the resulting state machine. The deep learning part of it never happens on the GPU of the DLSS end user, it's just running the net nvidia trained, which the article got right... and I'd be interested to see how much of the tensor core capacity is in actual use with features like this. I'll quit my rant here and fix the grammar and try to improve the paragraph a bit later, but I couldn't help myself, marketing speak and the current disaster that is the GPU market price that resulted from it angers me.

Cheers A Shortfall of Gravitas (other machine) (talk) 08:22, 4 April 2021 (UTC)

Article still contains inaccuracies, contradictory claims, and marketing language
Currently the second paragraph claims that "As of June 2021, this technology is available exclusively on GeForce RTX 20 and GeForce RTX 30 series of graphics cards.". Aside from being an obviously false claim, it seems to me to be a clear violation of MOS:PUFFERY. Either way this untrue claim has stood in the article since the creation when Hervegirod put it in. It's also contradicted later in the article. 62.248.185.4 (talk) 20:31, 7 May 2022 (UTC)

"Tensor cores", "Tensor Cores", or "tensor cores"?
How should it be spelled? "Tensor cores", "Tensor Cores", or "tensor cores"?

--Mortense (talk) 13:31, 25 November 2022 (UTC)

DLSS 1.9 to this day is complete conjecture
Neither Nvidia nor Remedy have ever confirmed a thing such as DLSS 1.9 existing. When Remedy implemented DLSS 1.0 into the game Control it surpassed the quality of similar implementations in other games.

After the release of DLSS 2.0 someone coined the previous implementation DLSS 1.9 which spread like wildfire.

But it was only coined because of it's qualitative similarity to 2.0. Which never excluded Remedy and Nvidia working together to provide a better implementation (and better machine learning training).

Which does not at all imply it being a different implementation or version of DLSS 1.0 at all. Only that Remedy's own TAA implementation and their implementation of DLSS 1.0 was an improvement.

But the same could have been said about the implementations of DLSS 1.0 in Metro Exodus and for example Battlefield V. Both of which had implementations so vastly differing in quality that one could imply they were different versions of DLSS 1.0 when they were not.

This self proclaimed fact of DLSS 1.9 being an actual differing iteration from DLSS 1.0 comes from the perception that DLSS 1.9 is/was the same (or almost the same) algorithm as DLSS 2.0 but ran on shader cores instead of shader cores.

This think is fueled by the presumption (and common disdain for Nvidia marketing and sales tactics) that Nvidia forced obsolescence by making DLSS 2.0 only work on RTX GPUs when it could have worked on all GPUs as shader-based implementation.

Which is completely unfounded, biased and complete speculation.

I believe much of this is based on a wishful perception by opponents of Nvidia and an unfortunate miswording or misunderstanding by Techspot.com who likely coined the term or at the very least picked it up in a more official journalistic capacity and therefor helped its propagation. (see https://www.techspot.com/article/1992-nvidia-dlss-2020/)

The only facts that remain are that only DLSS 1.0 and DLSS 2.0 officially ever existed as upscaling techniques and that within releases of games with DLSS 1.0 support we always had vast differentials in image quality.

These differences have never been proven to be anything but the quality of implementation and quality of game specific machine learning data.

Which as we know was one major factor in the different qualities of DLSS 1.0 implantations. The fact that DLSS 1.0 had to be trained on a per-game basis.

I therefor conclude that, until Nvidia ever officially states otherwise, DLSS 1.9 never existed.

Remedy's implementation of DLSS 1.0 simply outshone other implementations either by amount of effort to their own engine or/and to the amount of game specific machine-learned training that made people believe it was a different version entirely. Fnna509 (talk) 11:49, 13 August 2023 (UTC)

DLSS 3 is NOT Exclusive to Ada Lovelace NVIDIA GPUs.
DLSS 3 Has 2 main components: 1. Upscaling 2. Frame Generation While DLSS Frame Generation is indeed exclusive to Ada Lovelace NVIDIA GPUs, DLSS 3 (Subsequently, the third generation of DLSS) Upscaling isn't. 85.198.63.121 (talk) 16:07, 30 October 2023 (UTC)