With GDC ongoing, there's lots of info flying out from developers and hardware vendors and now AMD has given talks and more detail on FidelityFX Super Resolution 2.0. If you need a quick primer, see our original post from the initial announcement.
One of the main things is that it might be more difficult to integrate for developers. It requires developers provide it with more data points but AMD say it will be easier to integrate with games that "already have a temporal upscaling rendering path". They estimate if that rendering path exists it could be "as little as a few days" to integrate FSR 2.0.
Along with some updated screenshots they gave out a trickle of new info. Here's one of the new comparison shots, click it to enlarge it (warning - large file size):
It will support various quality modes just like FSR 1.0, here's what they've changed to for FSR 2.0:
FSR 2.0 Quality Mode |
Description | Scale factor | Input resolution |
Output resolution |
Quality | “Quality” mode provides similar or better than native image quality with a projected significant performance gain. |
1.5x per dimension (2.25x area scale)(67% screen resolution) |
1280 x 720 1706 x 960 2293 x 960 2560 x 1440 |
1920 x 1080 2560 x 1440 3440 x 1440 3840 x 2160 |
Balanced | “Balanced” mode offers an ideal compromise between image quality and projected performance gains. |
1.7x per dimension (2.89x area scale)(59% screen resolution) |
1129 x 635 1506 x 847 2024 x 847 2259 x 1270 |
1920 x 1080 2560 x 1440 3440 x 1440 3840 x 2160 |
Performance |
“Performance” mode provides image quality similar to native image quality with a projected major performance gain. |
2.0x per dimension (4x area scale)(50% screen resolution) |
960 x 540 1280 x 720 1720 x 720 1920 x 1080 |
1920 x 1080 2560 x 1440 3440 x 1440 3840 x 2160 |
As for hardware support, as they already announced it continues to be cross-vendor and cross-platform too. Since it's now an "advanced temporal upscaling solution" it will be a little more demanding to use.
Here's what they suggest as a starting-level (minimum) for FSR 2.0, which may change over time:
Target Upscaling Resolution | AMD Graphics Cards | NVIDIA® Graphics Cards |
4K |
Radeon™ RX 6700 XT |
GeForce RTX™ 3070 |
1440P |
Radeon™ RX 6600 |
GeForce RTX™ 3060 |
1080P |
Radeon™ RX 6500 XT |
GeForce® GTX 16 Series |
No release date has yet been set but games will start appearing with FSR 2.0 in "Q2 2022" and sometime around then it will see the wider open-source release of FSR 2.0 for everyone.
See their full talk about FSR 2.0 below:
![YouTube Thumbnail](https://www.gamingonlinux.com/cache/youtube_thumbs/6e4fbdc168f80f7cb14c6aa2889d2184.jpg)
Direct Link
Quoting: GuppyQuoting: JahimselfOn the screenshot, if you watch it full picture, and look the first face on the foreground, you can better compare native to other FSR. You can observe how it alterate the overall image.
The white bit with the "carbon fiber" texture? yeah I did wonder why the 'quality' seems to double the size of the pattern not of the other seems to do that.
Maybe something like a "fixed scale" post-process pixel shader? I mean, they might simply render a bitmap or algorithmic pattern on top of certain objects, maybe masked with a Gaussian high-pass filter to keep it from flattening and overpowering the underlying shape. That would mean that upscaling the resulting image simply upscales the pattern as well.
I'm not completely certain but I think the other faces/heads actually show the same exact pattern, unscaled but blurred or simply at a lower opacity. Anyway, I think this bears a relation to how constant width outline algorithms work these days, now that fixed function pipelines have been made completely redundant.
To further support my hypothesis, I see the pattern clearly in the "Balanced" and "Performance" reference images as well, maybe slightly larger and blurrier than in "Quality". Not quite sure it's a 50% difference between the extremes though.
Forgive my ignorance of proper terminology. I'm obviously out of my comfort zone. Now someone who actually knows something about 3D graphics should feel free to tell me I'm an idiot. It's fine, I can take it, and will be wearing my dunce cap for the rest of the day as usual.
To get back to the actual topic: The gradient banding on "Performance" and slightly jarring effects of sharpening bother me a bit. But I suppose "Quality" does give an illusion of native resolution quite convincingly, if you ignore the surfaces where textures get oversharpened to (over)compensate for the inevitable blur. I know we're not all equally sensitive to that particular pet peeve.
In "Balanced" mode, something a bit weird happens to the guy sitting in profile on the rooftop in the middle distance, against the sunlit cliff face. Amplified motion blur from the new temporal scaling algorithm?
But honestly, my trusty old 1080p TV screen just keeps telling FSR to get a real job. My GPU can still run everything I throw at it at this resolution, without ever raising a ruckus. Well, except for things it was never designed to do like RT I suppose.
Quoting: GuppyIt is called "dithering", and unfortunately when you upscale image it upscales too, independently on FSR or bilinear on any other.Quoting: JahimselfOn the screenshot, if you watch it full picture, and look the first face on the foreground, you can better compare native to other FSR. You can observe how it alterate the overall image.
The white bit with the "carbon fiber" texture? yeah I did wonder why the 'quality' seems to double the size of the pattern not of the other seems to do that.
Quoting: axredneckDoesn't look like dithering to me, more like a crosshatching pattern.Quoting: GuppyIt is called "dithering", and unfortunately when you upscale image it upscales too, independently on FSR or bilinear on any other.Quoting: JahimselfOn the screenshot, if you watch it full picture, and look the first face on the foreground, you can better compare native to other FSR. You can observe how it alterate the overall image.
The white bit with the "carbon fiber" texture? yeah I did wonder why the 'quality' seems to double the size of the pattern not of the other seems to do that.
See more from me