AMD Potentially Preparing to Announce FSR 2.0 at GDC 2022

[Uskompuf]

AMD is scheduled to hold an event discussing "Next-Generation Image Upscaling for Games" at the Game Developers Conference on March 23. The event only includes a brief description that "AMD will present some of the results of their research in the domain of next-generation image upscaling technology" but the developer of CapFrameX has recently claimed to see footage from FidelityFX Super Resolution (FSR) 2.0 so AMD may be preparing to announce the technology imminently.

The developer claims that FSR 2.0 switches to a temporal upscaling approach with optimized anti-aliasing that doesn't require AI acceleration unlike DLSS & XeSS meaning that it can work with GPUs from multiple vendors. The technology can also allegedly improve image quality beyond native resolution but we will need to wait for the official announcement and reviews before reaching any conclusions.



View at TechPowerUp Main Site | Source

 

[Deleted member 185088]

That's a good news, we need a universal solutions. Wonder why Microsoft doesn't push their upscaling more and make part of DX12, so all games can benefit from upscaling.

 

[Guwapo77]

This is outstanding news as I wasn't expecting an official announcement about this so soon. I need to know how this work for Cyberpunk 2077, I will dive back into game if I keep above 60 fps with ray tracing... Yeah I know, I got an AMD card. I'm already ready for next gen.

 

[ToTTenTranz]

Here are CapFrameX's tweets on the subject:

https://twitter.com/i/web/status/1502375332829188097

https://twitter.com/i/web/status/1503334843236687874

FSR Performance mode runs at 25% of the original resolution (50% per dimension).

25% of the original resolution is also what Unreal Engine 5's TSR uses by default. It's looking like a TSR-equivalent with the advantage of being cross-engine and whose open-source implementation is properly documented in the FidelityFX suite.

This could be huge for devices like the Steam Deck where FSR1.0 doesn't work great because of the low base resolution, for example.

This is outstanding news as I wasn't expecting an official announcement about this so soon. I need to know how this work for Cyberpunk 2077, I will dive back into game if I keep above 60 fps with ray tracing... Yeah I know, I got an AMD card. I'm already ready for next gen.

If it's a temporal upscaling method then it probably needs motion vectors to be implemented.
The good news is that games that use UE4/5 and/or DLSS already has those in place, so the development effort to implement FSR2.0 in such games should be small.
Cyberpunk 2077 is one of those cases, but we'd still need developer intervention regardless.

I'm also happening to be playing CP2077 at the moment on a 6900XT, and I have to say playing the game with a better upscaling/upsampling technology would be nice so I can up the raytracing settings.

Some more information about FSR 2.0. It looks like it will launch on Deathloop.

AMD FSR 2.0 'next-level temporal upscaling' officially launches Q2 2022, RSR launches March 17th - VideoCardz.com

Confirmed: AMD FSR 2.0 will be presented at GDC 2022 Next-gen FSR deliver higher image quality than 1.0. The information in this post is based on official slides, hence we will only report on what is confirmed thus far. AMD is announcing its FSR 2.0 technology on March 17th. This announcement...

videocardz.com

AMD seems to be claiming that "Performance Mode" that runs at 25% resolution has a similar quality to original, but with close to 2x the performance.

This is great on discrete GPUs, but on the Steam Deck this means the GPU can afford to render at only 640*400 which is a gamechanger considering the device's limited fillrate and memory bandwidth.

 

[Bomby569]

The "better then native" claims always makes me laugh

 

[ToTTenTranz]

The "better then native" claims always makes me laugh

TAAU at ~75% resolution does present better than native results at least on still image quality, though. It makes sense that temporal reconstruction that uses the previous 3-4 frames to resolve the current image presents a "better than native" result.

The question is whether or not there's too much or discernible ghosting, though.

 

[HisDivineOrder]

TAAU at ~75% resolution does present better than native results at least on still image quality, though. It makes sense that temporal reconstruction that uses the previous 3-4 frames to resolve the current image presents a "better than native" result.

The question is whether or not there's too much or discernible ghosting, though.

Ghosting inherently means it cannot be better than native.

 

[watzupken]

The "better then native" claims always makes me laugh

I think this is not impossible because an upscaling technology like DLSS can get rid of some image quality issues present on pure native resolution. However, there is no silver bullet to the problem because there will be trade offs when it comes to image quality. Moreover, we need to understand that upscaling technology‘s primary objective is to boost frame rates, and minimising image quality lost is likely secondary. And I feel that both DLSS and FSR have done a great job at this.

Ghosting inherently means it cannot be better than native.

Again, there are trade offs. Native image quality is not perfect either because you may see some shimmering which something like DLSS can remove. So “better than native” depends on which angle you are looking from.

 

[Bomby569]

I think this is not impossible because an upscaling technology like DLSS can get rid of some image quality issues present on pure native resolution. However, there is no silver bullet to the problem because there will be trade offs when it comes to image quality. Moreover, we need to understand that upscaling technology‘s primary objective is to boost frame rates, and minimising image quality lost is likely secondary. And I feel that both DLSS and FSR have done a great job at this.

Fanboys aside, has no one ever seen FSR pics were details just simply disapear? Even if you could improve something, there's hundred other places were you loose pixels.

 

[Punkenjoy]

I think the future of rendering is things like that. This is only the beginning. I think they will be able to do good with FSR 2.0 but I can't wait to see how they will address the ghosting issue.

DLSS do not use their neural network to create pixel from scratch, but it's being used to select witch pixel to re-use. I wonder if AMD will come up with a "normal" algorithm that will do the same and be able to identify and remove ghosting.

Back to my first point, 3d real time rendering was always about cheating to get results and anything that you can do to improve performance and/or image quality without requiring increased resources will make it's way into news games. Upscaler are here to stay, and they will continue to improve and at some point they will all the time deliver better than native performance and image quality. They might do in the future Variable upscaling a bit like VRS. An upscaling that will focus on rendering the part that would benefits more from a new frame and re-use previous frame from parts that are less visible.

As long as it's available to all and not locked into a single vendors, those technology will become standard.

This FSR 2.0 and XeSS probably announce the death of DLSS. Not because it's not good, but because it's closed to a single brand.

 

[nguyen]

I think the future of rendering is things like that. This is only the beginning. I think they will be able to do good with FSR 2.0 but I can't wait to see how they will address the ghosting issue.

DLSS do not use their neural network to create pixel from scratch, but it's being used to select witch pixel to re-use. I wonder if AMD will come up with a "normal" algorithm that will do the same and be able to identify and remove ghosting.

Back to my first point, 3d real time rendering was always about cheating to get results and anything that you can do to improve performance and/or image quality without requiring increased resources will make it's way into news games. Upscaler are here to stay, and they will continue to improve and at some point they will all the time deliver better than native performance and image quality. They might do in the future Variable upscaling a bit like VRS. An upscaling that will focus on rendering the part that would benefits more from a new frame and re-use previous frame from parts that are less visible.

As long as it's available to all and not locked into a single vendors, those technology will become standard.

This FSR 2.0 and XeSS probably announce the death of DLSS. Not because it's not good, but because it's closed to a single brand.

Let see how many games support FSR 2.0 and XeSS first, not sure game devs want to spend time implementing unknown upscalers vs tried and true DLSS, well unless AMD/Intel sponsoring those titles.

 

[Punkenjoy]

Let see how many games support FSR 2.0 and XeSS first, not sure game devs want to spend time implementing unknown upscalers vs tried and true DLSS, well unless AMD/Intel sponsoring those titles.

There is plenty of games that are not AMD sponsored that have FSR 1.0 just because anyone can grab the code and put it in their game since it's open source. We don't know yet about XeSS and FSR 2.0 but FSR 2.0 at least should be open source too.

As for DLSS, it's the same, Nvidia have deep pockets but it will be mostly in the future in Nvidia sponsored game.

 

[ToTTenTranz]

This FSR 2.0 and XeSS probably announce the death of DLSS. Not because it's not good, but because it's closed to a single brand.

One needs to wonder how Nvidia will justify all that die area dedicated to tensor cores if DLSS falls into disuse.
Though I guess they will push dev relationships as hard as they can to keep DLSS relevant, and in the case of Unreal Engine it's an automatic toggle apparently.

Let see how many games support FSR 2.0 and XeSS first, not sure game devs want to spend time implementing unknown upscalers vs tried and true DLSS, well unless AMD/Intel sponsoring those titles.

I bet most developers will prefer to adopt upsampling/reconstruction methods that they can implement transversally through PC and 9th-gen consoles.
Unless FSR2.0 is significantly inferior to DLSS (though we should expect it to be close to UE5's TSR), it's DLSS who might become something contained to Nvidia sponsored titles.

 

[rutra80]

The "better then native" claims always makes me laugh

When you'll see it you'll shit bricks.
Made me laugh too but then I got RTX and gotta admit that DLSS, DLAA, and DL scaling technology is here to stay.
Why are you able to take incredible photos with teeny weeny smartphones' image sensors? Because AI-combined shots from 3 tiny sensors are better than native medium sensor. Similar thing does DL with motion vectors and temporal magic.

As for ghosting, at 240Hz three-frames ghosting is better than native 60Hz.

Our brain does DL all the time, thanks to that we are often able to see more than our eyes' optics permit - it's pretty native.

 

[nguyen]

There is plenty of games that are not AMD sponsored that have FSR 1.0 just because anyone can grab the code and put it in their game since it's open source. We don't know yet about XeSS and FSR 2.0 but FSR 2.0 at least should be open source too.

As for DLSS, it's the same, Nvidia have deep pockets but it will be mostly in the future in Nvidia sponsored game.

AAA games nowaday either support both DLSS/FSR or none at all (like Elden Ring).

DLSS has been widely adopted already, kinda hard for any game dev not to know about it. FSR2.0 and XeSS on the other hand are still unknown upscalers.

One needs to wonder how Nvidia will justify all that die area dedicated to tensor cores if DLSS falls into disuse.
Though I guess they will push dev relationships as hard as they can to keep DLSS relevant, and in the case of Unreal Engine it's an automatic toggle apparently.

I bet most developers will prefer to adopt upsampling/reconstruction methods that they can implement transversally through PC and 9th-gen consoles.
Unless FSR2.0 is significantly inferior to DLSS (though we should expect it to be close to UE5's TSR), it's DLSS who might become something contained to Nvidia sponsored titles.

UE4.26 and UE5 game engines have DLSS plugin, no point for game devs not to click on the DLSS box unless they are paid not to do so (by AMD/Intel).

Most game devs should have experience with DLSS by now, while FSR2.0 or XeSS are unknown upscalers. Heck most current FSR1.0 implementations are unusable.

 

[Punkenjoy]

One needs to wonder how Nvidia will justify all that die area dedicated to tensor cores if DLSS falls into disuse.
Though I guess they will push dev relationships as hard as they can to keep DLSS relevant, and in the case of Unreal Engine it's an automatic toggle apparently.

Rumors are that Nvidia is splitting it's architecture to do one compute and one gaming. I wouldn't be surprised if Nvidia ditch the tensor core, or just don't add any of them and just keep them to continue to run DLSS. (or they could port DLSS to use more common compute.)

That should bring them more competitive on the power usage side as they won't have to power silicon that is not really used in a gaming scenario.

 

[Makaveli]

AAA games nowaday either support both DLSS/FSR or none at all (like Elden Ring).

DLSS has been widely adopted already, kinda hard for any game dev not to know about it. FSR2.0 and XeSS on the other hand are still unknown upscalers.



UE4.26 and UE5 game engines have DLSS plugin, no point for game devs not to click on the DLSS box unless they are paid not to do so (by AMD/Intel).

Most game devs should have experience with DLSS by now, while FSR2.0 or XeSS are unknown upscalers. Heck most current FSR1.0 implementations are unusable.

I think the point here is the "Unknown upscalers"

Are open source so the dev's can use them at will without additional cost.

 

[nguyen]

I think the point here is the "Unknown upscalers"

Are open source so the dev's can use them at will without additional cost.

DLSS cost nothing to the game studio either, and their games get promoted by Nvidia too

 

[Bomby569]

DLSS cost nothing to the game studio either, and their games get promoted by Nvidia too

apart from us the devs are the ones with the more to gain, imagine all the peasents with 650ti's that can now run CP77 and buy the game

 

[nguyen]

apart from us the devs are the ones with the more to gain, imagine all the peasents with 650ti's that can now run CP77 and buy the game

More like pirate the game .
But yeah I wouldn't buy AAA games when I have an old GPU, only to rely on upscalers to play at 1080p, that's just self tormenting.

 

[Guwapo77]

Here are CapFrameX's tweets on the subject:

https://twitter.com/i/web/status/1502375332829188097

https://twitter.com/i/web/status/1503334843236687874

FSR Performance mode runs at 25% of the original resolution (50% per dimension).

25% of the original resolution is also what Unreal Engine 5's TSR uses by default. It's looking like a TSR-equivalent with the advantage of being cross-engine and whose open-source implementation is properly documented in the FidelityFX suite.

This could be huge for devices like the Steam Deck where FSR1.0 doesn't work great because of the low base resolution, for example.



If it's a temporal upscaling method then it probably needs motion vectors to be implemented.
The good news is that games that use UE4/5 and/or DLSS already has those in place, so the development effort to implement FSR2.0 in such games should be small.
Cyberpunk 2077 is one of those cases, but we'd still need developer intervention regardless.

I'm also happening to be playing CP2077 at the moment on a 6900XT, and I have to say playing the game with a better upscaling/upsampling technology would be nice so I can up the raytracing settings.

Some more information about FSR 2.0. It looks like it will launch on Deathloop.

AMD FSR 2.0 'next-level temporal upscaling' officially launches Q2 2022, RSR launches March 17th - VideoCardz.com

Confirmed: AMD FSR 2.0 will be presented at GDC 2022 Next-gen FSR deliver higher image quality than 1.0. The information in this post is based on official slides, hence we will only report on what is confirmed thus far. AMD is announcing its FSR 2.0 technology on March 17th. This announcement...

videocardz.com

AMD seems to be claiming that "Performance Mode" that runs at 25% resolution has a similar quality to original, but with close to 2x the performance.

This is great on discrete GPUs, but on the Steam Deck this means the GPU can afford to render at only 640*400 which is a gamechanger considering the device's limited fillrate and memory bandwidth.

I am a stickler for quality modes. I underestimated just how nice raytracing is when implemented correctly, I should have gone Nvidia this round. Anyways, I'm only trying to play the game in 1440p with max ray tracing @ 60+ fps. I didn't jump to 4K because the GPUs weren't ready imo... I don't like turning down settings for framerates, I'd rather wait for the next GPUs to push the limit before playing. What you shared above does give hope I can play CP before the next gen cards are available.

 

[Minus Infinity]

The "better then native" claims always makes me laugh

Well I don't know about the upscaling that GPU makers use, but Adobe has an AI upscaling technology called super-resolution and I can attest as a photographer, it definitely increases IQ beyond the original. The results can be staggeringly good. It also corrects a lot of lens aberrations like distortion and can greatly enhamnce very fine detail. I have a 42MP camera and using super-resolution which upscales by 2x and then down scaling back to original resoltuion gives me files that are far more detailed than even the newer version of my camera that has 61MP.

I honestly could not believe what I was seeing. I will be reprocessing some of better landscape photos now.

So it might make you laugh now I guarantee it is possible.

Another example is noisy images. Topaz Denoise AI can actually restore features lost to noise because the AI knows what should be there.

AI image processing is already doing amazing things and it's it's still early days.

Now one way you can slightly improve IQ is to upscale an image (no AI) do some image processing and then resample back to native resolution. So I expect they mean if you ran FSR 2.0 without down-scaling, the resultant image could look better than native, but not by much.

 

[Camm]

Not a fan to the shift of everything being temporal. Faster paced games now can give me motion sickness.

FSR being spatial, whilst not as good IQ, at least gave an option that wasn't temporal based. Now it seems thats coming to an end. RIP.

 

[wolf]

I dont know why people bawk at the idea that at least some parts of the image can be better than native, as if Native is always some holy grail of perfection with no flaws or drawbacks whatsoever, totally unable to be improved upon, when spitting out the same final number of pixels. Even FSR 1.0 has managed this on rare occasion.

Seeing is believing.

 

[kapone32]

AAA games nowaday either support both DLSS/FSR or none at all (like Elden Ring).

DLSS has been widely adopted already, kinda hard for any game dev not to know about it. FSR2.0 and XeSS on the other hand are still unknown upscalers.



UE4.26 and UE5 game engines have DLSS plugin, no point for game devs not to click on the DLSS box unless they are paid not to do so (by AMD/Intel).

Most game devs should have experience with DLSS by now, while FSR2.0 or XeSS are unknown upscalers. Heck most current FSR1.0 implementations are unusable.

Will you ever say anything good about AMD's GPUs and software? BTW FSR allows you to play AAA Games at 1080P with Pascal GPUs. It is interesting that DLSS's expansion directly came right after FSR 1.0 launch. So you can actually thank AMD for your sentiment.