Nvidia ,Rtx2080ti,2080,2070, information thread. Reviews and prices September 14.

[crisium]

Page 35 paragraph 3:

"Whereas TAA renders at the final target resolution and then combines frames, subtracting detail, DLSS allows faster rendering at a lower input sample count, and then infers a result that at target resolution is similar quality to the TAA result, but with roughly half the shading work."

Is this what he means? Does that actually mean DLSS is running at a lower base resolution? Hmm something is fishy about that sentence...

 

[dlerious]

AIB unboxing from Jayz
https://www.youtube.com/watch?v=SlmK2rWGDyM

 

[Malogeek]

From the whitepaper, the 2070 TPCs and SMs are grouped into 3 GPCs as opposed to 6 GPCs for the 2080. I wonder how that will affect performance with GPC access to cache and memory controllers compared to TU104. For comparison, Pascal had 4 GPCs for 1080 and 3 for 1070.

 

[Muhammed]

No it is NOT an upscale.

It's a post process frame buffer AA technique that operates at the given resolution, using a DL network to properly smooth out the aliasing.

IT IS, how do you think it is providing close to 2X the performance of TAA?

 

[PeterScott]

Page 35 paragraph 3:

"Whereas TAA renders at the final target resolution and then combines frames, subtracting detail, DLSS allows faster rendering at a lower input sample count, and then infers a result that at target resolution is similar quality to the TAA result, but with roughly half the shading work."

Is this what he means? Does that actually mean DLSS is running at a lower base resolution? Hmm something is fishy about that sentence...

I think that is more about DLSS using less frames and less samples from the frames, not about there being any less native resolution.

If they are running at lower than native resolution, then they there will be some grief coming their way when people figure that out. Because there certainly haven't stated anything like that, and it makes the comparisons much more disingenuous.

If DLSS is really comparing 1800p + upscale to native 4K, then it just looks like a big cheat.

 

[Muhammed]

I think that is more about DLSS using less frames and less samples from the frames, not about there being any less native resolution.

Nope, here is the exact description from PCGH write up on the issue:

Nvidia indicates that DLSS uses a non-rocked resolution below the set one, which is supplemented by deep learning training with temporal component supersamples. You could say that this is a new form of upscaling that extrapolates not just existing information, but adds new ones. In addition, there should be DLSS 2X, a mode that works with the actual resolution and applies an equivalent of 64 × supersampling. Of course, this mode is slower than the native rendering.

In short, there are two modes of DLSS:
DLSS 1X: uses lower resolution, upscales it perfectly to look like native resolution +TAA
DLSS 2X: uses native resolution and super samples it at very low performance hit
http://www.pcgameshardware.de/Gefor...7862/Specials/Turing-Technik-Infos-1264537/3/

 

[Muhammed]

If they are running at lower than native resolution, then they there will be some grief coming their way when people figure that out.

Why would there by any grief? You get native res quality at significantly more fps. This is a killer deal.

 

[crisium]

^^Well if that's true, I'm still very intrigued. Depends on DLSS 2X performance hit.

 

[Hitman928]

Page 35 paragraph 3:

"Whereas TAA renders at the final target resolution and then combines frames, subtracting detail, DLSS allows faster rendering at a lower input sample count, and then infers a result that at target resolution is similar quality to the TAA result, but with roughly half the shading work."

Is this what he means? Does that actually mean DLSS is running at a lower base resolution? Hmm something is fishy about that sentence...

I'm not an expert by any means in graphics pipelines / rendering techniques, but my understanding is that TAA works by combining samples through time (frames) and then subtracting (ideally) any aliasing. DLSS basically works by comparing a high res version of the image to the rendered image and then uses the trained network to make the native res image look as close to possible as the high res version but still at the native resolution. So it's not the same as SSAA and won't look as good as SSAA or actually running at a higher native resolution, but it's kind of the same idea as SSAA. How good it looks will depend on how long they can train the network for and how quickly the tensor cores can do the inference.

With standard DLSS I would still expect some "softening" of geometry edges and maybe some blur here and there compared to running at a higher native res.

 

[PeterScott]

Why would there by any grief? You get native res quality at significantly more fps. This is a killer deal.

First, because they haven't been upfront with what they are doing. If it's checkerboard rendering instead of native, of course it's going to be faster. People still complain that the PS4 Pro isn't doing proper 4K because a lot of games use checkerboard rendering. It would have been much better to be honest from the start that this is what they are doing.

Second, because scaling techniques can look good in parts, it often has glitches and artifacts, that NVidia isn't showing us in marketing mode, but will be quickly be found by users, when the product ships.

 

[Malogeek]

Here's the info from the whitepaper on the 2 DLSS modes

The key to this result is the training process for DLSS, where it gets the opportunity to learn how to produce the desired output based on large numbers of super-high-quality examples. To train the network, we collect thousands of “ground truth” reference images rendered with the gold standard method for perfect image quality, 64x supersampling (64xSS). 64x supersampling means that instead of shading each pixel once, we shade at 64 different offsets within the pixel, and then combine the outputs, producing a resulting image with ideal detail and anti-aliasing quality. We also capture matching raw input images rendered normally. Next, we start training the DLSS network to match the 64xSS output frames, by going through each input, asking DLSS to produce an output, measuring the difference between its output and the 64xSS target, and adjusting the weights in the network based on the differences, through a process called back propagation. After many iterations, DLSS learns on its own to produce results that closely approximate the quality of 64xSS, while also learning to avoid the problems with blurring, disocclusion, and transparency that affect classical approaches like TAA.

In addition to the DLSS capability described above, which is the standard DLSS mode, we provide a second mode, called DLSS 2X. In this case, DLSS input is rendered at the final target resolution and then combined by a larger DLSS network to produce an output image that approaches the level of the 64x super sample rendering - a result that would be impossible to achieve in real time by any traditional means. Figure 23 shows DLSS 2X mode in operation, providing image quality very close to the reference 64x super-sampled image.

 

[Muhammed]

Second, because scaling techniques can look good in parts, it often has glitches and artifacts, that NVidia isn't showing us in marketing mode, but will be quickly be found by users, when the product ships.

They already shown it in the Infiltrator demo that was available to the public, people tore the demo apart looking for differences, they found none.

First, because they haven't been upfront with what they are doing.

They have been upfront about it from day one. Didn't you read any previews or interviews of the Turing announcement?

People still complain that the PS4 Pro isn't doing proper 4K because a lot of games use checkerboard rendering.

This is much higher quality than checkerboarding.

 

[Muhammed]

Here it is from NVIDIA's own description:

Whereas TAA renders at the final target resolution and then combines frames, subtracting detail, DLSS allows faster rendering at a lower input resolution, and then infers a result that at target resolution is similar quality to the TAA result, but with half the shading work.

https://devblogs.nvidia.com/nvidia-turing-architecture-in-depth/

 

[PeterScott]

They have been upfront about it from day one. Didn't you read any previews or interviews of the Turing announcement?

Show me any place NVidia said they were running at less than native resolution for DLSS. They just said it's 4K+ DLSS. They didn't say it is 1800P + DLSS scaled to 4K, or checkerboard + DLSS scaled to 4K.

NVidia just said it was 4K + DLSS.

 

[PeterScott]

Here it is from NVIDIA's own description:

https://devblogs.nvidia.com/nvidia-turing-architecture-in-depth/

Good catch. I stand corrected. But that just went up today.

And note that they didn't word it like that in the PDF.

I have been following this all along, because DLSS was the feature I was most interested in, and this was the first indication from NVida that I saw that mentioned they were running DLSS at lower resolution.

 

[Hitman928]

Here's the info from the whitepaper on the 2 DLSS modes

Ok, so I was right. They're rendering things at 64x SSAA and then asking the neural network to match the native res image to the super-scaled image.

DLSS 2x technique is a little less clear. Sounds like they're actually rendering the image at 64x resolution and then feeding that into the neural network and combining it with some other learning algorithm such that the final native image should look like it has 64x SSAA.

 

[Muhammed]

Ok, so I was right. They're rendering things at 64x SSAA and then asking the neural network to match the native res image to the super-scaled image.

No.

Here it is from NVIDIA's own description:

Whereas TAA renders at the final target resolution and then combines frames, subtracting detail, DLSS allows faster rendering at a lower input resolution, and then infers a result that at target resolution is similar quality to the TAA result, but with half the shading work.

https://devblogs.nvidia.com/nvidia-turing-architecture-in-depth/

So here is how DLSS works: NVIDIA super computer renders the images at 4K + 64X SSAA. Builds an AI training model around those perfect images. Then use that information to make the game run at less than 4K native resolution, but upscales it through AI to the quality of the 64X SSAA native res.

 

[PeterScott]

I also feel there is a bit of DLSS bait and switch. Those highly improved image quality vs TAA samples are for DLSS 2X.

We will have to wait for reviewers to get their hands on games that support DLSS to get the real story on the image quality differences between TAA/DLSS/DLSS 2X.

And to find out the performance hit for DLSS 2X.

 

[Muhammed]

New I also feel there is a bit of DLSS bait and switch. Those highly improved image quality vs TAA samples are for DLSS 2X.

The infiltrator demo used DLSS 1X. IQ was the same between the two.

 

[crisium]

Funny they changed "lower input resolution" to "lower input sample count".

They are definitely trying to make it discreet.

But if the results add up then that's fine by me. This is the checkerboarding option PC games need, but with better results. We will also have DLSS2X for the truly good stuff.

 

[Hitman928]

No.

Here it is from NVIDIA's own description:



https://devblogs.nvidia.com/nvidia-turing-architecture-in-depth/

So here is how DLSS works: NVIDIA super computer renders the images at 4K + 64X SSAA. Builds an AI training model around those perfect images. Then use that information to make the game run at less than 4K native resolution, but upscales it through AI to the quality of the 64X SSAA native res.

Ok, that's a new explanation. Wonder why they didn't include that in the white paper? This makes more sense for the performance uplift as TAA is not performance heavy so there was no way they could get 2x performance as compared to TAA, you wouldn't get 2x performance from just turning off AA entirely.

It will be interesting to see quality comparison in motion if they are starting at sub native resolution. I also am curious how it scales down in quality. The image difference between 720p and 1080p is a lot bigger than 1440p to 4K, IMO, despite the relative ratio being the same. What I'm getting at is I'm curious if it works really well at 4K where they can sub-scale from 1440p and it still looks really good versus trying it at 1080p where you would sub-scale from 720p and the visual difference would be too much for the trained network to still make it look good? I don't know, just musing here.

 

[Malogeek]

From my understanding, the DNN is making huge supersampled versions to determine the ground truth so that for particular scenes or geometry (?) it knows ahead of time what is the ideal subsample pixel location for a pixel when it hits a geomtry boundary that requires subsampling. What I don't understand is what the training consists of and how that data is then finally stored and provided to the driver.

With sometime like infiltrator demo, there's no interaction so you can sample each frame and it's identical each time. With a game it's all dynamic, you don't know ahead of time where the camera is and what it's facing. Do they have monkeys at Nvidia playing the game at 64x SSAA running around everywhere for days on end? Does the game developer have to provide a method of interaction with the game engine that the DNN can hook into and render all possible scenes and angles? That's an insane amount of data to analyze and then what exactly is provided to the driver for the game?

 

[PeterScott]

The infiltrator demo used DLSS 1X. IQ was the same between the two.

Same as TAA? That isn't saying much. It now looks like that static screen shots that were shown for the higher than TAA quality were DLSS 2X, which we didn't know existed previously. I believe this is alos from infiltrator. What does regular DLSS look like for this? This is why we need reviews and not cherry picked marketing:

 

[Hitman928]

Funny they changed "lower input resolution" to "lower input sample count".

They are definitely trying to make it discreet.

But if the results add up then that's fine by me. This is the checkerboarding option PC games need, but with better results. We will also have DLSS2X for the truly good stuff.

TAA combines multiple frames whereas DLSS doesn't, so by "default" it will user a lower input sample count. So, yea, by changing the wording in the white paper they're obscuring what the technique actually does. Like you said though, if it works, it works. It's like lossless memory compression, if you can get higher performance with no image quality loss, who's going to complain about that? Obviously I'll wait for independent reviews for the actual image quality comparison but the performance gain is potentially significant.

 

[Despoiler]

They already shown it in the Infiltrator demo that was available to the public, people tore the demo apart looking for differences, they found none.

It's far easier to do all sorts of very complex things when the output is known and repeatable ahead of time. That's what RTX demos have all been about right? Pre-baked RT and DLSS demos...woo woo marketing. Just like we found out that raytracing will be 30-60fps @ 1/4 resolution, DLSS is probably going to have its own set of limitations. Realtime[unpredictable] movement is a challenge for any frames they haven't trained which will be most of them if they are quoting "thousands". I'm curious if any temporal issues will come up where you get fed a version of the frame that doesn't match what is actually going on in game.