Alex from The Temple of Ao Fei
Qubits | Official account QbitAI
Friends familiar with graphics cards should know that Nvidia's DLSS can improve resolution and frame rate at the same time, making the game experience more silky.
But this pioneering AI rendering technology only supports new cards above the RTX20 series, so what about old card players?
No, the "savior" came.
At GDC2022, AMD unveiled the mystery of their new technology, FidelityFX Super Resolution 2.0 (FSR2.0).
AMD said the technology doesn't require specific machine learning hardware, not even machine learning algorithms, but instead uses hand-coded algorithms.
AMD believes that manual algorithms have more room for control to adapt to different game scenarios.
This means that to experience super resolution, you don't have to throw away your old card and replace it with an expensive new graphics card.
And this good thing?
Let's take a look at how well FSR 2.0 works.
FSR2.0 VS FSR1.0
FidelityFX Super Resolution is a technology released by AMD in 2021, benchmarking NVIDIA DLSS.
Although FSR1.0 has many advantages, it still has some obvious drawbacks.
It requires high-quality anti-aliasing source images, and this is a difficult problem to solve: games without anti-aliasing must also follow this rule when using FSR 1.0, which makes data integration more time-consuming.
Since FSR1.0 leverages functions to boost the resolution of the input image, when the source resolution is very low, there is not enough information to regenerate the details, and even the flickering of the picture and poor edge reconstruction can be seen, which are more obvious in the performance upgrade preset.
To achieve a technological breakthrough, AMD uses advanced time algorithms to reconstruct details from scratch, replacing spatial scaling with time scaling.
Finally, the new FSR 2.0 was born, which was not built on top of FSR 1.0 and did not take the path of AI acceleration. Compared with the "predecessors", the main improvements of FSR2.0 are:
Different inputs can be accepted, and if a jagged source image is imported, FSR2.0 can take advantage of anti-aliasing to optimize the output image quality.
Different image quality modes are available and support dynamic resolution scaling, which users can choose according to their needs.
No specific machine learning hardware is required, such as NVIDIA Tensor cores, Intel XMX units. Through the library, complete C++ and HLSL source code is available, as well as API documentation to support image fusion.
As long as the game is supported, you can get performance improvements on almost any graphics card, including NOT ONLY AMD's own graphics cards (including RX 6000, RX 5000, RX Vega, RX 400, Ryzen™ APU), but also support NVIDIA and Intel graphics cards.
It can be seen that when the "Deathloop" game turned on the light chase, FSR 2.0 was significantly better than FRS 1.0 (paying special attention to wall textures), sharper and clearer.
AMD YES?
AMD also acknowledges that some of the current FSR 2.0 optimizations need to be improved.
For example, while in all of AMD's examples, it doesn't run at more than 1.5 milliseconds, it doesn't do on low-end GPUs.
However, in the space of 1.5 milliseconds, FSR 2.0 did a lot of things.
It replaces a full time anti-aliasing channel by calculating motion vectors, reprojecting frames to eliminate jitter.
Also create "unobstructed masks" that compare one frame to the next, observing what moved and what didn't, to eliminate ghosting effects that lock in fine features where appropriate, such as nearly invisible stair edges and thin wires.
There are also preventing color drift, sharpening the entire image, etc.
AMD said it will make FSR 2.0 available to developers next quarter and will continue to open source it, followed by samples, APIs and plugins on the GPUOpen website.
In addition, AMD said that both Deathloop and Forspoken will use FSR 2.0.
While it's not exactly when Microsoft's Xbox game developers will take advantage of FSR 2.0, Xbox will also fully support the technology.
So this time, do you feel AMD yes?
Reference Links:
[1]https://alexewerlof.medium.com/my-guiding-principles-after-20-years-of-programming-a087dc55596c
[2]https://blog.daftcode.pl/hype-driven-development-3469fc2e9b22