AMD unveils new FidelityFX Super Resolution 3.0 details at GDC 2023

[Jimmy2x]

Forward-looking: During last year's Ryzen 7000 series launch, AMD announced updates to its FidelityFX Super Resolution upscaling solution. Team Red doubled down on its promise at this year's Game Developer Conference, presenting more details surrounding the benefits, challenges, and technologies related to FSR 3.0's development.

During its FidelityFX™ SDK presentation, AMD's Jason Lacroix treated attendees to a more detailed preview of the tool's upscaling features and capabilities. The next iteration of FidelityFX Super Resolution (FSR 3.0) will give users a two-fold increase in frame performance compared to the existing FSR 2.0 solution. Unfortunately, AMD didn't have any dates concerning the technology's release.

Team Red says it can already deliver on the claims using currently available solutions and a concept known as frame interpolation. Frame interpolation is used in up-sampling technologies to synthesize pixels and frames between generated frames or images. The technique is crucial in increasing upscaled frame-per-second performance without sacrificing overall image quality.

If successfully implemented, AMD engineers are confident they can achieve more than twice the framerate boost while achieving a one-to-one ratio of sample and interpolated pixels. The solution's successful implementation would also ensure no feedback loops in the generation process, so any artifacts or other inaccuracies would appear only once and not be factored into future frames.

The road to reaching the desired goal of doubling frame performance is no simple task. According to the presentation, several challenges remain before AMD can successfully implement FSR 3.0. Team Red did not shy away from discussing these challenges, highlighting potential issues related to color clamping techniques, non-linear motion interpolation, and increased interpolation requirements for postprocessing and user interface elements.

AMD's upscaling solution is a software-based, open-source technology that can be leveraged across multiple generations and brands of graphics hardware. The solution is fighting an uphill battle against Nvidia's hardware-based Deep Learning Super Sampling (DLSS) technology. DLSS 3.0, which has been available for adoption and use since the RTX 4000 series graphics cards launched in October of last year, relies on proprietary, hardware-based Tensor cores to support its upscaling solutions.

FidelityFX Super Resolution is a software-based, open-source technology that developers can leverage across multiple generations and brands of graphics hardware. The solution is fighting an uphill battle against Nvidia's hardware-based DLSS. Deep Learning Super Sampling 3.0, available since the RTX 4000 series graphics cards launched in October, relies on proprietary hardware-based Tensor cores to support upscaling.

While AMD's solution is more than a generation behind Nvidia's latest iteration, adherence to an open, software-based solution could be the differentiator for developers looking to adopt an upscaling methodology. The upcoming FSR 3.0 technology aims to provide developers and users with excellent overall performance, reduced latency, and a more streamlined integration path for new games and those already utilizing FSR 2.0. In contrast, Nvidia's DLSS 3.0 is exclusive to the latest lineup of 4000-series GPUs.

If AMD's FSR 3.0 continues the company's trend of supporting older, and even competing, hardware while delivering on promises of enhanced visual quality and frame delivery, then Team Red may be able to close the growing gap in the battle for upscaling technology adoption.

Permalink to story.

https://www.techspot.com/news/98075-amd-unveils-new-fidelityfx-super-resolution-30-details.html

 

[ScottSoapbox]

These fake double frame rates are going to do more to hurt game optimization than help playability.

 

[VariableSpike]

These fake double frame rates are going to do more to hurt game optimization than help playability.

Exactly, usually tech like this means devs take their foot off the gas more rather than it actually improving performance as they spend less time optimising

 

[TheRipper]

These fake double frame rates are going to do more to hurt game optimization than help playability.

Agreed. It will just become a crutch because investors want to cut corners and performance will tank until we're at the point where they will introduce 2x frame generation. Rinse, repeat.

 

[emmzo]

Older cards profiting from this means they shot themselves in the foot when they`ll want to sell new gen. That`s why sly Huang keeps it proprietary and restricted to last gen. It doesn`t mean AMD cares more. It just means they don`t have the resources to fully work on it and leaving it open source, it entices other to work for them for free.

 

[neeyik]

These fake double frame rates are going to do more to hurt game optimization than help playability.

Not necessarily. Frame generation requires upscaling to be used, so to get the best performance with the system, the main rendering still needs to be good at, say, 1080p. Targeting something like 30 fps will result in an awful FSR 3.0 experience, whereas 60 or 90 will be far nicer. And it's also much easier to optimize for 1080p than it is for 4K.

 

[ScottSoapbox]

Not necessarily. Frame generation requires upscaling to be used, so to get the best performance with the system, the main rendering still needs to be good at, say, 1080p. Targeting something like 30 fps will result in an awful FSR 3.0 experience, whereas 60 or 90 will be far nicer. And it's also much easier to optimize for 1080p than it is for 4K.

Regardless of target resolution, the most likely impact of this is that devs will say the game can play at 60 FPS for mainstream settings and hardware when it really is at 30 FPS with doubling.

Consoles have raised the standard to 60 FPS (even though they aren’t actually always hitting it) which is a benefit for gamers. This largely undoes that.

 

[nnguy2]

Not necessarily. Frame generation requires upscaling to be used, so to get the best performance with the system, the main rendering still needs to be good at, say, 1080p. Targeting something like 30 fps will result in an awful FSR 3.0 experience, whereas 60 or 90 will be far nicer. And it's also much easier to optimize for 1080p than it is for 4K.

In two years 8k gaming will mean 1080p upscaled to 8k

 

[mbk34]

In two years 8k gaming will mean 1080p upscaled to 8k

My poor eyes will still be stuck at 720p

 

[neeyik]

Regardless of target resolution, the most likely impact of this is that devs will say the game can play at 60 FPS for mainstream settings and hardware when it really is at 30 FPS with doubling.

Consoles have raised the standard to 60 FPS (even though they aren’t actually always hitting it) which is a benefit for gamers. This largely undoes that.

A significant portion of PC games that offer upscaling are console ports or those made in parallel development. If the console version is achieving 60 fps at 1080p and the same at higher resolutions with upscaling, then there's no reason to instantly assume that the PC version will be running at half the frame rate.

In two years 8k gaming will mean 1080p upscaled to 8k

If you couldn't tell, just be looking at the images or playing the game, that upscaling (with or without frame generation) was used, would it actually matter?

 

[nnguy2]

If you couldn't tell, just be looking at the images or playing the game, that upscaling (with or without frame generation) was used, would it actually matter?

That's a big if. I can see difference now on an image upconverted to 4k and more than likely I'll see it at 8k.

My analogy is audio. A number of people can't tell the difference between compressed audio vs uncompressed audio format. But just because someone can't hear the difference doesn't mean I can't.

I'm cursed with good discerning eyesight and hearing.

 

[neeyik]

That's a big if. I can see difference now on an image upconverted to 4k and more than likely I'll see it at 8k.

Upscaling, be it via DNNs or shader routines, currently aren't good enough to hide the fact it's been used. Two years, of course, is probably going to be too short a period of time for AMD, Intel, and Nvidia to improve their algorithms (and hardware) to get to the point where it's not noticeable at 1440p/4K (especially given the current GPU market).

But there will come a time when it will all be good enough -- by that time, though, games are likely to be using it by default. One may never be able to discern anything, simply because there'll be no reference point without upscaling.

To me, it's akin to compression in GPUs. When its use first started to appear with textures, the application of it was extremely visible, though the performance gain was useful. Now, compression is used across a broad spectrum of parts in a GPU and it's completely transparent to the end user.

 

[Bamda]

Thanks, Jensen. sigh!

 

[Hodor]

FSR 8 will be so good, you'll play Tetris and it will upscale the picture to look like you're playing Call of Duty.