News AMD Universal Upscaling With Radeon Super Resolution, FSR 2.0 Incoming

[Admin]

AMD Software Adrenalin Edition version 22.3.1 adds support for RSR upscaling which works with virtually all games. AMD also previews the upcoming FSR 2.0, and other driver features.

AMD Universal Upscaling With Radeon Super Resolution, FSR 2.0 Incoming : Read more

 

[hotaru.hino]

Couple of comments;

• Where's the screenshots at their actual output resolution? It's kind of hard to compare a 4K image when it's shrunk down to 970x574
• If you do plan on doing a comparison, please use NVIDIA's Image Scaler in NVCP instead of DLSS since it's more of an apples-to-apples comparison. Unless I missed the memo FSR 2.0 switched to temporal-based reconstruction.

 

[salgado18]

Couple of comments;

• Where's the screenshots at their actual output resolution? It's kind of hard to compare a 4K image when it's shrunk down to 970x574
• If you do plan on doing a comparison, please use NVIDIA's Image Scaler in NVCP instead of DLSS since it's more of an apples-to-apples comparison. Unless I missed the memo FSR 2.0 switched to temporal-based reconstruction.

• There is a button bellow the expanded image that lets you see it in the original size.
• I believe it should have both, but between the two, I'd prefer it to be compared to DLSS, since it is one of the big features of the RTX series.

 

[cryoburner]

Where things get interesting is when we look at where FSR is going with FSR 2.0, and consider what that might mean for RSR. If AMD can apply FSR 1.0 at a driver level and get good results, and if RSR has similar performance requirements but generates better image quality, there's no reason AMD can't upgrade RSR in the future to the better FSR implementations.

I'm not so sure they would be able to easily update RSR to use FSR 2.0, since the new algorithm apparently requires more data than just the completed image. FSR 2.0 appears to require access to the depth buffer, containing the distance to each pixel, as well as a motion buffer, containing the motion of each pixel, to be able to predict where pixels rendered in one frame will appear in subsequent frames, making it more like DLSS. And I don't get the impression that a driver-level upscaler would be able to accurately determine how to properly utilize that data without developer intervention.

AMD doesn't have tensor cores on its current RX 6000 or previous RX 5000 series GPUs, and as the minority provider of graphics hardware, it wants to make FSR work with the widest range of GPUs possible.

Of course there's also the consoles running AMD GPUs, which will likely make heavy use of FSR. And really, if you consider that there are close to 170 million PS4 and Xbox One consoles that many developers are still supporting, and already over 30 million PS5 and Series X/S consoles out there, along with around 20 million AMD GPUs being used on Steam, it's probably not accurate to say that AMD is a "minority provider of graphics hardware", at least in the combined PC/console gaming space.

Judging by Steam's reported monthly users and percentage of those with Nvidia hardware, we can extrapolate that there's likely close to 100 million Nvidia GPUs being used for gaming on Steam. Though only around 30 million of those are RTX 20/30 cards that can utilize DLSS. So there's somewhere around 200 million AMD GPUs being used for gaming, and probably a roughly similar amount of Nvidia GPUs once we add the Nintendo Switch, but fewer than 10% of those devices can currently utilize DLSS, whereas a majority of the hardware should be able to utilize FSR. And that's before getting into Intel's GPUs and the many mobile devices that should also be compatible with FSR.

 

[hotaru.hino]

• There is a button bellow the expanded image that lets you see it in the original size.
• I believe it should have both, but between the two, I'd prefer it to be compared to DLSS, since it is one of the big features of the RTX series.

• Maybe my browser wasn't working or something because I clicked on it before and it gave me the preview image
• Well, the lack of including the image upscaler in NVCP needs addressing. I don't feel like this feature gets enough of a spotlight and people simply assume NVIDIA's answer to FSR is only DLSS.

 

[JarredWaltonGPU]

Couple of comments;

• Where's the screenshots at their actual output resolution? It's kind of hard to compare a 4K image when it's shrunk down to 970x574
• If you do plan on doing a comparison, please use NVIDIA's Image Scaler in NVCP instead of DLSS since it's more of an apples-to-apples comparison. Unless I missed the memo FSR 2.0 switched to temporal-based reconstruction.

You missed the memo: "FSR 2.0 will bring improved image quality compared to FSR 1.0 at all upscaling quality presets. This will come in part from a switch from spatial upscaling to temporal upscaling with "optimized" anti-aliasing features. Spatial upscaling means the only data used for upscaling a frame comes from the frame itself. Temporal upscaling in contrast can use data from the current frame as well as previous frames. Interestingly, Nvidia's DLSS 1.0 was a spatial upscaling solution, and DLSS 2.0 switched to temporal upscaling."

 

[JarredWaltonGPU]

I'm not so sure they would be able to easily update RSR to use FSR 2.0, since the new algorithm apparently requires more data than just the completed image. FSR 2.0 appears to require access to the depth buffer, containing the distance to each pixel, as well as a motion buffer, containing the motion of each pixel, to be able to predict where pixels rendered in one frame will appear in subsequent frames, making it more like DLSS. And I don't get the impression that a driver-level upscaler would be able to accurately determine how to properly utilize that data without developer intervention.

Of course there's also the consoles running AMD GPUs, which will likely make heavy use of FSR. And really, if you consider that there are close to 170 million PS4 and Xbox One consoles that many developers are still supporting, and already over 30 million PS5 and Series X/S consoles out there, along with around 20 million AMD GPUs being used on Steam, it's probably not accurate to say that AMD is a "minority provider of graphics hardware", at least in the combined PC/console gaming space.

Judging by Steam's reported monthly users and percentage of those with Nvidia hardware, we can extrapolate that there's likely close to 100 million Nvidia GPUs being used for gaming on Steam. Though only around 30 million of those are RTX 20/30 cards that can utilize DLSS. So there's somewhere around 200 million AMD GPUs being used for gaming, and probably a roughly similar amount of Nvidia GPUs once we add the Nintendo Switch, but fewer than 10% of those devices can currently utilize DLSS, whereas a majority of the hardware should be able to utilize FSR. And that's before getting into Intel's GPUs and the many mobile devices that should also be compatible with FSR.

True about AMD console hardware... and yet, most latest gen console games aren't using FSR on consoles, and previous gen console games aren't getting it either. MS and Sony seem to have their own upscaling solutions that are preferred instead, or maybe the devs are just rolling their own. Plus, saying 170 million consoles use AMD hardware only works if you include the previous generation stuff that uses effectively RX 470-580 GPUs. Anyway, FSR can run on the older consoles as well as Intel and Nvidia solutions. But that's mostly because it's a shader-based upscaling solution right now.

As for RSR 2.0 or whatever, it's possible AMD can just do the current and previous frame buffers and not worry about depth and motion buffers. We'll have to wait and find out exactly how FSR 2.0 works, but I don't expect it to do all the same temporal stuff DLSS is currently doing. AMD's stated goal is improved performance first, , probably hardware support second, and image quality third (my words, not theirs). A good temporal upscaling algorithm may not need the complexity of DLSS to work well.

 

[hotaru.hino]

You missed the memo: "FSR 2.0 will bring improved image quality compared to FSR 1.0 at all upscaling quality presets. This will come in part from a switch from spatial upscaling to temporal upscaling with "optimized" anti-aliasing features. Spatial upscaling means the only data used for upscaling a frame comes from the frame itself. Temporal upscaling in contrast can use data from the current frame as well as previous frames. Interestingly, Nvidia's DLSS 1.0 was a spatial upscaling solution, and DLSS 2.0 switched to temporal upscaling."

So basically the ball is in NVIDIA's court to make DLSS a driver wide feature. Or add another thing to the NVCP upscaler.

 

[JarredWaltonGPU]

So basically the ball is in NVIDIA's court to make DLSS a driver wide feature. Or add another thing to the NVCP upscaler.

Nvidia already has a driver upscaling solution, though it's not necessarily as good as FSR. I'd need to do more in-depth testing to try to determine that. I don't think DLSS will ever be fully integrated into the drivers as an "apply everywhere" type of thing, but maybe Nvidia will prove me wrong.

 

[cryoburner]

Plus, saying 170 million consoles use AMD hardware only works if you include the previous generation stuff that uses effectively RX 470-580 GPUs

Having only RX 480-class hardware seems like all the more reason to make use of a good upscaler. And as prior-gen console users transition to current-gen consoles, the number of active AMD GPUs in those devices is likely to remain relatively similar.

We'll have to wait and find out exactly how FSR 2.0 works, but I don't expect it to do all the same temporal stuff DLSS is currently doing. AMD's stated goal is improved performance first, , probably hardware support second, and image quality third (my words, not theirs). A good temporal upscaling algorithm may not need the complexity of DLSS to work well.

There's a slide from AMD showing that FSR 2.0 is utilizing frame buffer and motion vector data, and it seems likely that those are necessary to see most of the improvements. I'm sure there will be notable differences to how DLSS works behind the scenes, but from an implementation standpoint, the process will likely be relatively similar for developers utilizing it.

 

[hotaru.hino]

There's a slide from AMD showing that FSR 2.0 is utilizing frame buffer and motion vector data, and it seems likely that those are necessary to see most of the improvements. I'm sure there will be notable differences to how DLSS works behind the scenes, but from an implementation standpoint, the process will likely be relatively similar for developers utilizing it.

DLSS 2.0 works on the same principle, but it remains to be scene if running the reconstruction part using tensor cores with a generic AI model or shader units with an algorithm is better in the end. DLSS does have the advantage though that specific games can be augmented by giving the AI some training data and running that instead.

At least that's how I'm reading it from publicized material.