AMD’s gaming resolution patent answers many of the questions we had about the red team’s upcoming FidelityFX Super Resolution (FSR) … and in fact gives us a lot more to do. Whenever AMD is ready to talk about its potential response to DLSS, that is.
We have been pouring over the filing of patents published today (via Videocardz), although it was originally filed in November 2019. From what we can distinguish from the patent of the speakers, and from one and another that is often confused between Jacob and me, this implementation of the Super Resolution function of AMD should be an agnostic platform. in any game and does not necessarily need to run on a specific GPU.
If this is how AMD's DLSS-a-like feature works, it will have a much wider function than Nvidia's more restrictive technology. Although formed by game, rather than a "fully learned environment", DLSS could end up being more accurate in games for which it is specifically coded. However that The patent suggests that its approach may make the AMD version quite impressive in itself.
The more we hear about AMD’s FSR, the more it looks like the super-sampling equivalent of the FreeSync Vs. G-Sync. AMD’s technology is more open and works very well, while Nvidia’s is a closed design, but potentially a little more effective.
In the end, if AMD FSR is good enough, it will be a much more useful and widely used technology. And that’s pretty exciting.
However, we return to the specific details of this patent filing. What seems to differentiate FSR from DLSS is in the way it uses several larger-scale networks, both linear and nonlinear, to try to preserve important visual data from the original image when it explodes. But also in this sense, it aims to "super resolve images efficiently and widely." To us, this seems to mean not having to form a neural network in a specific game before it can be used.
The patent details the basic flow of this Super Resolution game such as taking a low resolution image (which can be quickly represented by the hardware in question) and then uses two deep learning networks, one linear and one another nonlinear, to reduce the sample in several different images. These differently sampled images allow FSR to remove a large number of different functions (for example, colors, objects, and curves) from a low-resolution starting point and then use it to create a more high-resolution image. detailed and accurate at the end.
"The combination of linear and nonlinear magnification," says the patent, "facilitates both the preservation of color and large-scale features (large objects and shapes that are more easily perceived by the human eye). of the image from the linear magnification as well as the preservation of the finer functions (for example, curved functions and features that are not easily perceived in low resolution) of the image from the rise to nonlinear scale.
"Linear operations use only input data, while nonlinear operations use both input data and other data (that is, non – input data) to increase input data. "Nonlinear functions make it easier to accurately determine complex features (for example, curves) of a more efficient image than nonlinear functions (for example, convolution operations)."
The multiple images sampled below are combined to create the pixels in the high-resolution image using another "pixel mix" stage. This involves using multiple transposed images to create blocks of pixels again by nine, per low-resolution pixel, using eight offset images, and one non-offset image of the images sampled down.
After this pixel blending process, some more cleaning operations are performed before creating a final high resolution image.
There seem to be a lot of things to do for each frame, but then it always seems like that when we look at what we expect a graphics card to do when we throw games at them.
This is just the simplified process, but the patent details the fact that the super resolution feature can use any or all of these elements, in any combination. This could suggest that, when used on different types of devices, it can be adapted to the hardware on offer. It is said to be used in "a computer, a gaming device, a handheld device, a decoder, a television, a mobile phone, or a tablet."
Nor does it necessarily have to run on a GPU, although this does give it the parallelism needed to make linear and nonlinear portions at the same time. "When hardware does not support parallel processing," says the patent, "linear upload processing and nonlinear upload processing are not performed in parallel."
He even claims that super resolution could run on software, although it would inevitably be much slower and would probably negate the benefits.
Still, it sounds pretty exciting, especially if it can really work, well, anything. If it can work effectively with old graphics cards, or even potentially with an Nvidia GPU, it could be an incredibly powerful tool. However, it will depend on how exactly FSR is activated.
If it’s an Adrenalin controller switch on the PC, it’s hard to see how Nvidia would enable it, though there’s a possibility it could be included in games or game engines, such as some of FidelityFX’s current features of AMD.
Maybe they could be both. There may be an agnostic version of the game that lives on the AMD graphics driver and another that is enabled for each game.
As I said at the beginning of this patent plan, this post leaves a lot of questions to be answered. Fingers crossed, we'll know sooner: Computex will arrive early next month, with a keynote speech by Dr. Su, and the virtual E3 is still very close. If there are rumors in June about a version of FidelityFX Super Resolution, these are the two best bets for its presentation.