AMD will talk about its "next-generation image upscaling technology" for games at the upcoming GDC 2022 conference later this month. The method is based on AMD's ongoing research, so it could be a new implementation of AMD's existing FidelityFX Super Resolution (FSR), though exactly what's being upgraded isn't yet clear.
"In this session AMD will present some of the results of their research in the domain of next-generation image upscaling technology, and how such technology can be applied to games to improve the gaming experience," reads a description of the Next-Generation Image Upscaling for Games session at the GDC 2022 website (kudos to VideoCardz for the tip). The event will take place on Wednesday, March 23, 2022.
AMD's existing FSR technology is a set of open-source upscaling algorithms designed to deliver improved performance at higher resolutions by upscaling a lower resolution source (e.g., upscaling 2560×1440 to 3840×2160, etc.). Unlike Nvidia's deep learning super sampling (DLSS), AMD's current implementation of FSR does not use any AI/DL-based image analysis to improve image quality, which is why the technology works on virtually all GPUs available today, including those from Intel and Nvidia.
But AMD's FSR in its current implementation is not perfect, and while it delivers on its promise of improved performance, in many cases the rendering quality of AMD's algorithm falls short. This is particularly true at higher upscaling factors, and there's little question that Nvidia's DLSS produces better results. That's why AMD is working on its next-generation image upscaling technology.
… Can Only Get Better
The company has not pre-announced details about its next-generation image quality improving methods, but we can speculate.
First up, since we are dealing with an upscaling technology for games that is ultimately designed to improve performance, we are talking about a real-time rendering algorithm that does not involve too expensive techniques that harm performance (like the multi-pass algorithms used by the movie industry).
Most high-quality image upscaling techniques introduced in the recent years (e.g., by Nvidia, Samsung, Sony, and LG) use various kinds of AI/DL analysis to recognize contrast patterns common in game-world/real-world scenes and then apply an upscaling algorithm (filter) that best fits a particular tile from a performance and quality point of view. Another way to improve performance while producing fine image quality is to implement separate algorithms to process the moving and still parts of an image, which again often uses some form of AI/ML-enabled analysis.
But while AI/ML-enabled upscaling methods work rather well right out-of-box with some input from the driver, the best way to implement them is to collaborate with game developers and implement a set of particular algorithms for a particular title. Movies don't care if it takes a bit longer to get a good upscaling result, but in a real-time game potentially running at more than 60 frames per second, complex techniques may not work as well.
There are several potential areas where AMD could look to improve on FSR. FSR currently uses spatial upscaling, which is something Nvidia attempted to do with DLSS 1.0. Nvidia eventually reworked the algorithm to use a temporal upscaler, combined with machine learning, and ultimately DLSS 2.0 has proven to be far better than the initial implementation.
Part of the reason Nvidia can get away with the complex DLSS calculations is that it equipped its RTX series GPUs with Tensor cores. These can perform a lot of less precise math very quickly, using FP16 instead of FP32 calculations. AMD's GPUs don't have an equivalent to Nvidia's Tensor cores, so there's a question of whether this next-generation upscaling technology is destined to work with current and previous generation GPUs, or perhaps it's for AMD's future RDNA 3 architecture.
At this point, we don't know if AMD's next-generation upscaling technology uses AI/DL analysis to boost performance and quality. Perhaps AMD has implemented hardware support for high-quality upscale algorithms used by programs like Adobe's Premier Pro or Blackmagic Design's DaVinci Resolve in its next-generation GPUs, but knowing GPU development trends, we would put our money on something more intelligent and less special-purpose.
AMD has traditionally tried to make algorithms that run on a wide set of graphics hardware, rather than locking the algorithm behind specialized hardware. Then again, Intel is also working on XeSS for its Arc GPUs, with a fallback method that will run on other non-Arc chips. If Intel can make XeSS work on GPUs from the past couple of generations, AMD might be attempting something similar.
We'll find out more about whatever AMD is working on with its image upscaling later this month at GDC 2022.
Considering the comparison in the Digital Trends' article, FSR does have some catching up to do in terms of quality. Though it's a very small sample set.