Support NeoGAF

[Negotiator]

Dude, I'm not surprised by your disingenuous behaviour. I remember your comments and you did the same thing in our argument about PS2 vs Xbox when you tried to present facts in such a way to make the inferior console superior. You wanted to tell people that the PS2 hardware could render the same effects in software and outperform the more powerful console in polycounts even though it was RAM limited and you couldnt even grasp that you need memory even to be able to use higher polycounts. Even developers who made games on all platforms said that Xbox had the upper hand when it came to hardware and polycounts, whereas the PS2 has the upperhand in other areas (fillrate) but you still thought Sony's console was superior. I had to link internal reports done by sony to show people the real poly counts of PS2 games and the numbers were nowhere near what you claimed and that finally shut you up.

PS2's exotic philosophy is a total antithesis to nVidia's philosophy.

Brute force (eDRAM insane 2560-bit bus @ 48GB/s) vs efficiency (hardwired circuitry, compression etc).

PS2 was capable of emulating fur shaders (SoTC), but the framerate was abysmal:

Shadow of the Colossus Making of (PS2 / PS3) ... Playstation 2 / Playstation 3

Shadow of the Colossus - the making of in english and with pictures

selmiak.bplaced.net

I'm not sure if XBOX's GeForce 3 GPU was capable of rendering fur shaders, but I vividly remember this GeForce 4 demo:



nVidia continues the efficiency philosophy with DLSS... brute force/rasterization is no more (even AMD will stop chasing that train with FSR4 aka rebranded PSSR).

I'm curious to see if PS6 will have something like DLSS4 + MFG, because I don't expect a huge bump in CPU/GPU specs (could be a sizeable jump if they adopt chiplets instead of a monolithic APU with low yields, we'll see).

 

[Corporal.Hicks]

PS2's exotic philosophy is a total antithesis to nVidia's philosophy.

Brute force (eDRAM insane 2560-bit bus @ 48GB/s) vs efficiency (hardwired circuitry, compression etc).

PS2 was capable of emulating fur shaders (SoTC), but the framerate was abysmal:

Shadow of the Colossus Making of (PS2 / PS3) ... Playstation 2 / Playstation 3

Shadow of the Colossus - the making of in english and with pictures

selmiak.bplaced.net

I'm not sure if XBOX's GeForce 3 GPU was capable of rendering fur shaders, but I vividly remember this GeForce 4 demo:



nVidia continues the efficiency philosophy with DLSS... brute force/rasterization is no more (even AMD will stop chasing that train with FSR4 aka rebranded PSSR).

I'm curious to see if PS6 will have something like DLSS4 + MFG, because I don't expect a huge bump in CPU/GPU specs (could be a sizeable jump if they adopt chiplets instead of a monolithic APU with low yields, we'll see).

Developers on the PS2 could render great looking effects in software (let's say comparable to xbox), but with performance penalty, like your example with fur rendering. Xbox could render these effects more easily thanks to dedicated pixel & vertex shaders and most importantly developers had enough RAM budget to use these shader effects on much bigger scale.

As for DLSS, it was a very clever idea, and without it I don't think Nvidia would push RTX technology as soon. The RTX 2080ti could run RT games well, but only with DLSS. Even RTX50 series cards will require to use DLSS if people want to get high refreshrate experience in the most demanding RT/PT games. Personally, I see no problem with this, given how well the technology works. I was not sure I would like nvidia FG, but it also works very well (at least x2 FG, as x3 or x4 is a bit too extreme).

Mark Cerny is going to give developers what they want. I'm sure the PS6 hardware will be build around AI (PSSR, ML powered FG, ML ray reconstruction).

 

[FireFly]

Okay fine, I'm happy to draw a line under you starting our discussion in this thread like below and thumbing up Corporal.Hicks childish verbal insults if you actual want to discuss.

Well, apart from using jargon, what I said only applies to you if you were actually arguing in bad faith.

Regardless of render output resolution and ML AI algorithm my view is no normal view of the windmill in those screenshots with that scene camera should have the contrast or detail shown

That's fine, but the purpose of DLSS is not to make an image more natural or "real". It is to get closer to the 16xSSAA "ground truth" image, that reflects how the developers intended the scene to look. So in this case the relevant question is whether the windmill has that level of detail in the SSAA anti-aliased image, and I suspect the answer is yes, based on the other distant scene elements.

 

[PaintTinJr]

Well, apart from using jargon, what I said only applies to you if you were actually arguing in bad faith.

Just own it and move on - don't do the bad faith thing you yourself floated wrongly about a legit point of analysis.

That's fine, but the purpose of DLSS is not to make an image more natural or "real". It is to get closer to the 16xSSAA "ground truth" image, that reflects how the developers intended the scene to look. So in this case the relevant question is whether the windmill has that level of detail in the SSAA anti-aliased image, and I suspect the answer is yes, based on the other distant scene elements.

Okay that's fine but DLSS is now copying the PSSR/Sony Bravia XR algorithm/patent with it's more intelligent handling of in scene data via it new buzzword 'Transformers', so by its own standards, XeSS' or FSR3.1's and before, yeah it's doing great....but that's old hat, much like its efforts with fur in R&C.

On DLSS 4 it should be much closer to PSSR by attempting to "shallow fake" the fur, rather than just reproduce standard heavy duty real-time rendering, so taking a ML AI agnostic view, I'm still disinterested in any of those efforts on the windmill, pixel peeping or not.

 

[ArtHands]

I wonder how many years will Sony take to also switch to Transformer model. 6 or 7 years maybe?

 

[3liteDragon]

I wonder how many years will Sony take to also switch to Transformer model. 6 or 7 years maybe?

I mean given the fact that PS6 is prolly gonna be out in 2028, we probably might see a switch during the early PS6 gen.

 

[PaintTinJr]

I wonder how many years will Sony take to also switch to Transformer model. 6 or 7 years maybe?

Go read Cerny's patent and listen to his recent technical video, or search the marketing explanations of how the Bravia X1 chips or newer XR processors work.

Nvidia are just using their own buzzword for what PSSR is already doing more intelligently. PSSR starting with a full-sized native image with holes to lower the pixel count is the best algorithm.

Training PSSR with games designed with it in mind will clear up these early issues. The only question is whether it will be on Pro games this gen or PS6 games next-gen.

 

[Negotiator]

Go read Cerny's patent and listen to his recent technical video, or search the marketing explanations of how the Bravia X1 chips or newer XR processors work.

Nvidia are just using their own buzzword for what PSSR is already doing more intelligently. PSSR starting with a full-sized native image with holes to lower the pixel count is the best algorithm.

Training PSSR with games designed with it in mind will clear up these early issues. The only question is whether it will be on Pro games this gen or PS6 games next-gen.

So PSSR is already superior from the get-go?

Because so far it's strongly reminiscent of DLSS1...

 

[Bojji]

So PSSR is already superior from the get-go?

Because so far it's strongly reminiscent of DLSS1...

I don't know people are discussing anything seriously with this poster. He lives in Sony fantasy world.

 

[Gaiff]

Go read Cerny's patent and listen to his recent technical video, or search the marketing explanations of how the Bravia X1 chips or newer XR processors work.

Nvidia are just using their own buzzword for what PSSR is already doing more intelligently. PSSR starting with a full-sized native image with holes to lower the pixel count is the best algorithm.

Training PSSR with games designed with it in mind will clear up these early issues. The only question is whether it will be on Pro games this gen or PS6 games next-gen.

Time to stop this. You spouted a lot of bullshit regarding PSSR and got humbled. You called out DF, claiming they couldn’t pixel count PSSR and it turns out they had no problem. They had the b-roll footage in their hands, but somehow, you knew better than them despite not having seen it and you crowned it above DLSS before even seeing it. I also recall Fafalda schooling you regarding this patent that wasn’t even PSSR. You even went as far as saying Rift Apart on the Pro is a better experience than a 4090-powered PC, once again, before getting your hands on the game.

This was a monumental L for you and you need to quit that.

 

[Topher]

So PSSR is already superior from the get-go?

Because so far it's strongly reminiscent of DLSS1...

I never used DLSS 1, but from what I understand DLSS 1 was a complete shitshow. PSSR is not.

 

[Negotiator]

I never used DLSS 1, but from what I understand DLSS 1 was a complete shitshow. PSSR is not.

Silent Hill 2 says hi.

 

[Topher]

Silent Hill 2 says hi.

I'm not going to do the list war thing, but you do know that there are other games that have had a lot better results with PSSR than SH2, right?

I'm not suggesting PSSR implementation has been perfect, but the results have been mixed at worst.

 

[PaintTinJr]

Time to stop this. You spouted a lot of bullshit regarding PSSR and got humbled. You called out DF, claiming they couldn’t pixel count PSSR and it turns out they had no problem. They had the b-roll footage in their hands, but somehow, you knew better than them despite not having seen it and you crowned it above DLSS before even seeing it. I also recall Fafalda schooling you regarding this patent that wasn’t even PSSR. You even went as far as saying Rift Apart on the Pro is a better experience than a 4090-powered PC, once again, before getting your hands on the game.

This was a monumental L for you and you need to quit that.

Are you saying Cerny doesn't know his own PSSR methodology, like he explained in his most recent video, and that aligns with his patent?

 

[PaintTinJr]

So PSSR is already superior from the get-go?

Because so far it's strongly reminiscent of DLSS1...

The training and implementation in games that weren't built with it in mind certainly aren't better. But the methodology was already superior, and the fur handling in R&C is the poster child for that shallow fake ML AI upscaling.

 

[Zathalus]

Go read Cerny's patent and listen to his recent technical video, or search the marketing explanations of how the Bravia X1 chips or newer XR processors work.

Nvidia are just using their own buzzword for what PSSR is already doing more intelligently. PSSR starting with a full-sized native image with holes to lower the pixel count is the best algorithm.

Training PSSR with games designed with it in mind will clear up these early issues. The only question is whether it will be on Pro games this gen or PS6 games next-gen.

Nvidia buzzwords? They didn’t develop the transformer architecture, that was mostly Google. Visual Transformers are a newer variant that are used in DALL-E and Stable diffusion for example.

 

[PaintTinJr]

Nvidia buzzwords? They didn’t develop the transformer architecture, that was mostly Google. Visual Transformers are a newer variant that are used in DALL-E and Stable diffusion for example.

You make it virtually impossible not to use the words 'man' and 'straw' when you don't read what I write verbosely. I made no mention of the hardware which is what you are referencing, I'm talking about the software terminology they've used, clearly after listening to Cerny's video and then did the terminology buzzword ' CNN' old, 'Transformer' new slight of hand.

Actually comprehend the methodologies each describe of DLSS4 and PSSR that get used and then tell me the big difference between DLSS3 and DLSS4 ML AI upscaling isn't already happening in the PSSR algorithm.

So whether they call it old CNNs vs new Super-water-in-to-wine-transformers it still amounts to the same thing of using CNN's in a layered way with intelligent processing of what objects the algorithm locates in the image,

 

[Zathalus]

You make it virtually impossible not to use the words 'man' and 'straw' when you don't read what I write verbosely. I made no mention of the hardware which is what you are referencing, I'm talking about the software terminology they've used, clearly after listening to Cerny's video and then did the terminology buzzword ' CNN' old, 'Transformer' new slight of hand.

Actually comprehend the methodologies each describe of DLSS4 and PSSR that get used and then tell me the big difference between DLSS3 and DLSS4 ML AI upscaling isn't already happening in the PSSR algorithm.

So whether they call it old CNNs vs new Super-water-in-to-wine-transformers it still amounts to the same thing of using CNN's in a layered way with intelligent processing of what objects the algorithm locates in the image,

I made zero mention of any hardware? I’m simply pointing out that transformers, more specifically visual transformers, is not a Nvidia buzzword. It is a well understood alternative to CNNs that is widely used in visual computing today. Visual transformers are not the same thing as a CNN.

As for what is happening in DLSS4 and how it compares to PSSR or DLSS3, I can’t tell you. Because nobody outside of Nvidia knows that information. I’ve certainly made no claim that visual transformers are going to be miles better vs a regular CNN, as I’m not going to pretend to know how each would be different for a temporal up scaler.

 

[Buggy Loop]

Nvidia buzzwords? They didn’t develop the transformer architecture, that was mostly Google. Visual Transformers are a newer variant that are used in DALL-E and Stable diffusion for example.

No but dude… look up Cerny’s secret sauce, totally not buzzwords for my plastic boxes, Nvidia is buzzwords.. even though they were first with CNN upscaling, but Cerny is legit.

 

[UnrealEck]

The multi-billion dollar company isn't going to fuck you bro.

 

[PaintTinJr]

I made zero mention of any hardware? I’m simply pointing out that transformers, more specifically visual transformers, is not a Nvidia buzzword. It is a well understood alternative to CNNs that is widely used in visual computing today. Visual transformers are not the same thing as a CNN.

As for what is happening in DLSS4 and how it compares to PSSR or DLSS3, I can’t tell you. Because nobody outside of Nvidia knows that information. I’ve certainly made no claim that visual transformers are going to be miles better vs a regular CNN, as I’m not going to pretend to know how each would be different for a temporal up scaler.

There's lots of information on both techniques available, and although their use of the term "transformer" would imply that DLSS4's upscaling uses full self-attention mechanism to derive global context of frame data per frame to provide a holistic processing approach across each frame.

The reality is that the difference in TOPS per cycle between a RTX 2060 and RTX 5090 is just a single digit multiplier or a very small two digit multiplier and the algorithm on the RTX 50xx card still needs to execute in 1ms, which isn't enough processing for self-attention - to generate per pixel pair relevance scores, etc - but would be enough for sparse attention, which is not holistically different from PSSR's use of analysing global image context via generatiing smaller mips or focusing on image features with layered CNNs.

In the end it is a lot of abstract terminology to describe semantically similar ML AI processing that is constrained by 1ms processing and the memory bandwidth or interconnect bandwidth to tensor cores available in that 1ms.

So in context, unless DLSS4 is somehow able to do the full self-attention mechanism per frame part in less than 1ms, rather than sparse attention - and do it on old RTX 2060 too - then the use of the word "transformer" is a bit disingenuous in a versus CNN slide IMO and is being exploited like a buzzword.

 

[Zathalus]

There's lots of information on both techniques available, and although their use of the term "transformer" would imply that DLSS4's upscaling uses full self-attention mechanism to derive global context of frame data per frame to provide a holistic processing approach across each frame.

The reality is that the difference in TOPS per cycle between a RTX 2060 and RTX 5090 is just a single digit multiplier or a very small two digit multiplier and the algorithm on the RTX 50xx card still needs to execute in 1ms, which isn't enough processing for self-attention - to generate per pixel pair relevance scores, etc - but would be enough for sparse attention, which is not holistically different from PSSR's use of analysing global image context via generatiing smaller mips or focusing on image features with layered CNNs.

In the end it is a lot of abstract terminology to describe semantically similar ML AI processing that is constrained by 1ms processing and the memory bandwidth or interconnect bandwidth to tensor cores available in that 1ms.

So in context, unless DLSS4 is somehow able to do the full self-attention mechanism per frame part in less than 1ms, rather than sparse attention - and do it on old RTX 2060 too - then the use of the word "transformer" is a bit disingenuous in a versus CNN slide IMO and is being exploited like a buzzword.

You’re overestimating the capabilities of the 2060 there. It takes over 2.5ms for the algorithm to upscale a 1080p image to 4K on that card. It’s also been confirmed that the newer DLSS4 has a slight performance penalty vs DLSS3, so probably close to 3ms on a 2060. Regardless of this speculation, the proof is in the results, so I guess we will know in a few weeks what improvements, and potential regressions, are in this newer DLSS.

 

[PaintTinJr]

You’re overestimating the capabilities of the 2060 there. It takes over 2.5ms for the algorithm to upscale a 1080p image to 4K on that card. It’s also been confirmed that the newer DLSS4 has a slight performance penalty vs DLSS3, so probably close to 3ms on a 2060. Regardless of this speculation, the proof is in the results, so I guess we will know in a few weeks what improvements, and potential regressions, are in this newer DLSS.

I don't think I made any claim about the 2060's ability to do DLSS4 in a timescale, and in my original post in the thread commented on how the disparity between the 50xx and previous series meant the runtime of DLSS4 wouldn't be a free lunch for older series owners, but your comment is consistent with what I was getting at, and despite DLSS not have the TOPs per cycle to achieve self attention used without MFG, the huge 33ms latency would give enough TOPs per native frame to use self-attention for MFG, which makes a lot of sense considering a full frame context analysis and holistic scaling per frame and interframe would be most needed to minimise instability and artefacts caused by MFG.

 

[kikonawa]

Could Switch 2 use any of this shit?

You may keep your expectations low with nintendo hardware mate

 

[BlackTron]

I'm in love with this argument that insists PSSR is better, with a clause that we might not see the proof until sometime into PS6 gen. That does leave you a pretty wide out, doesn't it.

 

[PaintTinJr]

I'm in love with this argument that insists PSSR is better, with a clause that we might not see the proof until sometime into PS6 gen. That does leave you a pretty wide out, doesn't it.

The reason is quite simple. PSSR's algorithm actually renders native pixels with hole regions to facilitate scaling down the workload to generate the source native data. So even in the worst possible situation, the worst case is that 3/4 of the pixels are wrong and 1/4 are perfect native pixels of the upscaled target resolution.

Where as in DLSS' case every pixel could be wrong in a frame, because they are rendered complete at the low resolution and then everyone of them is replaced with an inferred pixel. And because PSSR has guaranteed correct pixels in its data set it can then use the predicted data to infer what those pixels would be - if they had been holes too - allowing it to re-entrantly inference the accuracy of the data it predicted and alter it.

As for the waiting for it to prove itself, it turns out one of the biggest aspects of quality change from DLSS3 to DLSS4 is actually training with more advanced neural networks, so it would seem the training and the resources needed to compute the training are a significant part of a solution, which can either be done well and, slow and cheap, or fast and expensive.

 

[BlackTron]

The reason is quite simple. PSSR's algorithm actually renders native pixels with hole regions to facilitate scaling down the workload to generate the source native data. So even in the worst possible situation, the worst case is that 3/4 of the pixels are wrong and 1/4 are perfect native pixels of the upscaled target resolution.

Where as in DLSS' case every pixel could be wrong in a frame, because they are rendered complete at the low resolution and then everyone of them is replaced with an inferred pixel. And because PSSR has guaranteed correct pixels in its data set it can then use the predicted data to infer what those pixels would be - if they had been holes too - allowing it to re-entrantly inference the accuracy of the data it predicted and alter it.

As for the waiting for it to prove itself, it turns out one of the biggest aspects of quality change from DLSS3 to DLSS4 is actually training with more advanced neural networks, so it would seem the training and the resources needed to compute the training are a significant part of a solution, which can either be done well and, slow and cheap, or fast and expensive.

I'm gonna be honest, I stopped reading and started laughing about halfway into the next sentence shortly after "The reason is quite simple"

 

[PaintTinJr]

I'm gonna be honest, I stopped reading and started laughing about halfway into the next sentence shortly after "The reason is quite simple"

I think IGN signups are still available if you think your contribution efforts would be a better fit there.

 

[BlackTron]

I think IGN signups are still available if you think your contribution efforts would be a better fit there.

The mere fact that you think the above was a simple explanation is enough to disregard it. As has been said earlier in the thread no one is taking you seriously.

 

[Tazzu]

I have a question, DLSS is not very good at 1080p on a regular sized screen, will this be better? What about on laptop sized screens?

 

[PaintTinJr]

The mere fact that you think the above was a simple explanation is enough to disregard it. As has been said earlier in the thread no one is taking you seriously.

Simple enough for anyone with the pre-requiste knowledge to participate in discussing the technology in a meaningful way, yes.

So maybe it is your issue that this part of the discussion it too much for you, but it is interesting to see you hold strong opinions about DLSS versus other techniques by your comments and reactions, yet are admitting doing so out of brand loyalty rather than any understanding of how they each compare....interesting.

 

[BlackTron]

Simple enough for anyone with the pre-requiste knowledge to participate in discussing the technology in a meaningful way, yes.

So maybe it is your issue that this part of the discussion it too much for you, but it is interesting to see you hold strong opinions about DLSS versus other techniques by your comments and reactions, yet are admitting doing so out of brand loyalty rather than any understanding of how they each compare....interesting.

I don't even remember how many different cards found their way into my old mobo's dirty PCI slot...don't tell me about loyalty

Edit: Whoops, AGP slot. No, wait...it was PCI too, at the same time. Carry on

 

[moniker]

This caught my eye.

The AI parallax infills has improved so much, that the 3 fake frames looked better than 4000-series 1 fake frame. That's an accomplishment, less artifacts despite more fake frames

I believe this is mdrejhon but I'm not sure if he is active here anymore.

 

[3liteDragon]

This caught my eye.



I believe this is mdrejhon but I'm not sure if he is active here anymore.

So this is something that’ll be on the 4000 series cards as well for single FG since it’s a DLSS improvement?