A few days ago, AMD released a new supersampling anti-aliasing technology FSR 2.0, RSR, of which the former is more advanced, better effect, front-end benchmark NVIDIA DLSS, Intel XeSS.
At GDC 2022 Game Developers Conference, AMD unveiled detailed details of FSR 2.0 technology, including features, principles, game support, hardware support, quality mode, picture quality, performance, and more.
1. Characteristics
The main features of FSR 2.0 are eightfold:
-Time algorithm: The original algorithm was spatial, and after changing to a time algorithm, the picture quality is basically similar to or better than the original picture.
- Anti-aliasing: Replaces any TAA anti-aliasing directly within the game.
- High image quality: Better than FSR 1.0 in all quality grades and resolutions, and provides multiple quality modes with dynamic resolution scaling.
- No machine learning required: No specific hardware unit is required, more platforms and hardware are supported, and it can be flexibly controlled for different scenarios and easier to optimize.
- Cross-platform: In addition to AMD's own graphics card, it also supports NVIDIA, Intel products, and also supports Xbox series hosts.
- Open Source: As in the original era, the source code is publicly available on the AMD GPUOpen website, under the MIT license.
2. Principle
AMD believes that machine learning is not necessary for high-quality image scaling. Typically, real-time scaling based on machine learning uses only model learning to determine how to combine historical sampling to generate a scaled picture without generating new features based on scene objects.
AMD has researched, developed, and optimized a series of advanced algorithms that can also achieve this effect, which is FSR 2.0, and has more advantages than machine learning solutions, including more flexible control of different scenarios, better optimization capabilities, and support for more hardware and platforms.
FSR 2.0 requires three buffer frames of rendering resolution, namely depth buffer, motion buffer, and color buffer, and the final screen is output after integration through a simple and flexible API.
If the game already supports time-scaled render paths, it will be easier to support FSR 2.0.
Because of the replacement of TAA anti-aliasing, any post-processing effect that requires anti-aliasing needs to be scaled later.
Any post-processing effects that require deep buffering need to be scaled in advance.
3. Game support
For different types of games, FSR 2.0 is deployed differently.
First of all, games that already support DLSS 2.x are the easiest to implement, and the fastest takes less than 3 days.
Second, UE4/EU5 engine games that have integrated the FSR 2.0 plugin.
Third, games that support rendering and display resolution unlocking have the priority of TAA.
Finally, the normal game, which takes the longest time, takes at least 4 weeks.
4. Hardware support
FSR 2.0 has a wide range of hardware uses, including AMD and competitive graphics, as well as Xbox consoles.
However, the overall hardware requirements are higher than those of the original generation, and the scaling resolution of different targets is different.
The minimum graphics card requirements for each grade given by AMD are as follows:
4K: RX 6700 XT, RX 5700, RTX 3070, RTX 2070 or higher
2K: RX 6600, RX 6500, RX Vega Series, RTX 3060, RTX 2060, GTX 1080 or higher
1080p: RX 6500 XT, RX 590, GTX 16 Series, GTX 1070 or higher
Intel's did not mention, after all, has not been released, according to the above requirements, can only meet the 2K, 1080p target resolution.
5. Quality mode
FSR 2.0 supports 4 different quality modes with different zoom factors, image quality and performance.
Quality:
The image quality is similar or better than native, the performance improvement is not much, the length and height are scaled by 1.5 times, the overall scale is 2.25 times, and the input resolution is 67% of the output resolution.
Balanced mode:
Taking into account the improvement of image quality and performance, the length and height are scaled by 1.7 times, the overall scale is 2.89 times, and the input resolution is 59% of the output resolution.
Performance:
The image quality is similar to the native, the performance is greatly improved, the length and height are scaled by 2.0 times, the overall scale is 4 times, and the input resolution is 50% of the output resolution.
Extreme Mode (Ultra):
Newly added, the performance improvement is the greatest, the picture quality is still very close to the native, the length and height are scaled by 3 times, the overall scale is 9 times, and the input resolution is 33% of the output resolution.
6. Performance
Previously, taking DeathLoop as an example, it was claimed that the frame rate of the RX 6900 XT was 53FPS when the 4K native resolution, light tracking was turned on, ultra-high image quality, and SAM was turned on, and after the RSR 2.0 performance mode was turned on, the frame rate could soar to 101FPS.
This time no more frame rate data is given, but the maximum latency is highlighted.
In quality mode, the target resolution of the top graphics card 4K does not exceed 1.1ms, and the target resolution of the mainstream graphics card 1080p is less than 0.6ms.
If it is a performance mode, the above two indicators are less than 1ms and 0.5ms, respectively.
7. Picture quality
The screenshots are all from Death Loop.
The following two sets of original figures, namely native, quality mode, balanced mode, performance mode, find the difference. The original figure 4K click here to download
8. Landing
FSR 2.0 premiere support game is DeathLoop, and the second is Forspoken by Square Enix's studio Luminous Productions, which will be released in October.
AMD will soon announce FSR 2.0 developer resources on GitHub, including:
Develop APIs, API documentation, complete C++/HLSL repositories, DX12 instances (Vulkan or add them later), UE 4.26/4.27 engine plugins, Xbox GDKX instances.