Arm Unveils "Accuracy Super Resolution" Based on AMD FSR 2

[Vayra86]

Why on earth would you base anything of FSR...

Because Nvidia is pretty damn slow getting every game its DLSS pass.

Not that AMD is fast, but at least you can just toss it in there.

 

[InVasMani]

Yeah makes plenty of sense for mobile/portable displays in general.

 

[lemonadesoda]

It’s always better to have the graphics pipeline (whatever method) doing the upscaling, than to output low resolution and have the TV do the upscaling.

 

[Minus Infinity]

Why on earth would you base anything of FSR...

Point has been made a thousand times before: it's open source.

If you have the skill you can work on the code free of charge and implement your owned improved FSR. Try that with DLSS.

It could be just marketing. People without any technical knowledge could think that upscaling takes a 720p resolution image and creates a REAL 1080p image. Just remembering those TV shows where they can enhance a low quality pixelated image to a high resolution incredibly detailed image. If viewers of those Tv shows can believe this, then they can surely believe that upscaling will offer them trully higher resolution, perfectly detailed image.

There could be some truth in this. Probably an upscaled image from a 720p displayed on a 1080p monitor to be somewhat better than displaying 720p resolution on the same monitor without any upscaling algorithm messing with the final image. And I am talking about FSR here, not DLSS that does have some advantages.

In the end the typical user might not know how to change a resolution or might not want to change the resolution. There are psychological reasons. Me, about 10-12? years ago I was playing Borderlands on a 32'' 1080p TV with the resolution set at 720p, 30-40fps and enjoying it(9800GT or HD4870 main + GT620 I think for the physx). Others wouldn't enjoy it. Even the thought that the resolution is lower than what the monitor can provide, could lead them to order a much faster GPU/CPU/system.

There are AI based upscalers that can produce far higher quality than native res. I use upscaling in my phtoography a lot. I photograph birds and you are always cropping hard often reducing image size 75%. I have compared output from an upscaled 12MP image to a native 24MP image using Topaz Gigapixel AI and It was virtually impossible to tell the difference. If the original source material is good quality the "AI" upscaled version will be veryn good. In more extreme upscaling tests, one photographer showed images from his 24MP camera upscaled to 61MP of his other camera and again it was hard to tell which was native.

There are no non-AI upscalers I know can do anything like that. All the proven methods like Lanczos etc look bad when upscaling x 50% or more for realistic images that aren't just simple shapes etc.

AMD is admitting defeat, AI upscaling is coming, probably in FSR 4.

 

[AusWolf]

Point has been made a thousand times before: it's open source.

If you have the skill you can work on the code free of charge and implement your owned improved FSR. Try that with DLSS.


There are AI based upscalers that can produce far higher quality than native res. I use upscaling in my phtoography a lot. I photograph birds and you are always cropping hard often reducing image size 75%. I have compared output from an upscaled 12MP image to a native 24MP image using Topaz Gigapixel AI and It was virtually impossible to tell the difference. If the original source material is good quality the "AI" upscaled version will be veryn good. In more extreme upscaling tests, one photographer showed images from his 24MP camera upscaled to 61MP of his other camera and again it was hard to tell which was native.

There are no non-AI upscalers I know can do anything like that. All the proven methods like Lanczos etc look bad when upscaling x 50% or more for realistic images that aren't just simple shapes etc.

AMD is admitting defeat, AI upscaling is coming, probably in FSR 4.

It's not defeat. The only reasons FSR doesn't use AI (just yet) is 1. because AMD didn't want to restrict it to RDNA 3 and lock out all the previous generations of GPUs that don't have AI cores and 2. AMD doesn't have the experience with AI that Nvidia has. The current version of FSR is and always has been a stopgap with something better coming. The time it takes for gamers to upgrade to GPUs with AI cores gives AMD time to develop their AI upscaler. No rush there (yet).

Other than that, I agree.

 

[john_]

AMD is admitting defeat, AI upscaling is coming, probably in FSR 4.

They where probably doing what they did with FreeSync. Offered a solution that can be implemented everywhere with zero cost. And they won. Everyone is using their solution. Bringing AI in FSR 4 is just the next step, not a defeat. As we have FreeSync and FreeSync Premium in TVs/monitors, we will have non AI FSR and AI FSR in graphics. FSR 2.x/3.x will keep being implemented for years to come as the easiest option that can be used even will old graphics cards and FSR 4, let's hope it will be using the NPU in CPU's, not just AMD's CPU's, any CPU with an NPU. That way NPU's will make sense even to simple gamers not caring for the other uses of those NPUs. Nvidia of course will retain the advantage with DLSS and will keep finding new ways to use those tensor cores in it's GPUs, but thanks to NPUs in CPUs we might get solutions closer to those offered from Nvidia. We need that to avoid ending up in a market where Nvidia GPUs will be the only choice.

 

[TumbleGeorge]

Bringing AI

AI didn't exist.

 

[Tropick]

Colossal W for raspberry pi gamers

 

[tsunami2311]

Yah more Upscalers.

I rather Drop my detail setting before I ever use Upscaler again gain FPS. only then I would use one if that dont work.

Such Thing would be great if say you have 1080p monitor or what ever and you CAN NOT get you FPS target, it will certainly look better, with such upscale then running say 720p on monitor it self . Atlest on still image for most part, but once it moving hit PQ obvious to me and they all suffer from that , it just some deal with it better, and all them ugly too me in that regard.

To those that dont notice or dont care kudo's

 

[Minus Infinity]

It's not defeat. The only reasons FSR doesn't use AI (just yet) is 1. because AMD didn't want to restrict it to RDNA 3 and lock out all the previous generations of GPUs that don't have AI cores and 2. AMD doesn't have the experience with AI that Nvidia has. The current version of FSR is and always has been a stopgap with something better coming. The time it takes for gamers to upgrade to GPUs with AI cores gives AMD time to develop their AI upscaler. No rush there (yet).

Other than that, I agree.

Sure they'll use the term defeat, but call it what you like, they know they are behind. Also their hardware lacks tensor cores to handle the matrix math for AI as efficiently, I'll bet if they had an AI upscaler it would be slower anyway.

RDNA3 could have offered AI upscaler and old hardware could still use FSR2, so why gimp your newer cards?

 

[AusWolf]

RDNA3 could have offered AI upscaler and old hardware could still use FSR2, so why gimp your newer cards?

Probably because of the second reason: they're behind Nvidia on this tech, and it's not ready yet.

 

[Chrispy_]

Upscaling itself takes gpu power, on weak mobile gpu's this is relevant, games on switch using fsr 2 tend to use a sinplified version for that reason. I will say they probably took a workload that made the difference more noticable

Yeah, I just didn't realise it was that intensive.

Given that FSR 1.5X is slower than native on low-end hardware, it seems like a godawful fit.