A few days ago, we covered the news of AMD potentially working on a second generation version of FidelityFX Super Resolution that would get see an announcement at this year’s Game Developers Conference. AMD has a sponsored session planned for March 23rd where the company will talk about “next-gen upscaling technology” for an hour and present the results/breakthrough it has made through research.
Around the same time as the session’s announcement, the developer of CapFrameX came forward and claimed to have seen a working demo of “FSR 2.0” citing incredible performance that would be even better than native without any need for AI. More importantly, though, the tweet mentioning all this stated that FSR 2.0 would switch to temporal upscaling from the spatial upscaling that powers FSR currently, putting it on-par with how DLSS (and soon XeSS) works.
People quickly put two-and-two together and figured out that FSR 2.0 (no pun intended) would headline this upscaling event AMD was holding at GDC 2022. While this was only a rumor at the time with only the CapFrameX tweet to go by, three days later, we have confirmation that FSR 2.0 is, in fact, real and launching very soon with an early announcement.
FSR 2.0 at GDC 2022
Videocardz managed to get hold of the slides that AMD will use to unveil FSR 2.0 at GDC 2022 and they not only confirm the existence of FSR 2.0 in the first place but also give us a slight idea about the performance. I say slight because Videocardz did not share actual comparison slides they had access to, rather only the key highlights about the upscaling tech to preserve some secrecy.
Firstly, FSR 2.0 will actually be announced on March 17th instead of on March 23rd when AMD is holding their next-gen upscaling technology session. Perhaps, we will get to know a little bit more about FSR 2.0 at the session. Moreover, Radeon Super Resolution (RSR) that was first leaked back in December of 2021, will also see an announcement on March 17th, at last. RSR is like FSR 1.0 but built on the driver level instead, and will work across an ever larger variety of games since it does not need to be implemented on a per-game basis. Moreover, RSR will work on RX 5000+ GPUs which means the RDNA architecture is required for this.
Now, let’s go over the currently-known details of FSR 2.0 and how it differs from the current FSR. Right away, the most important bit here is the switch to temporal upscaling. FSR uses spatial upscaling to boost the FPS in games, in contrast to DLSS which relies on temporal upscaling. Temporal upscaling is objectively better, especially with motion/moving objects, which, I don’t know if you’ve noticed, happens a lot in games.
Now equipped with temporal upscaling, this new version of FSR would offer significantly better performance compared to FSR based on spatial upscaling. FSR 2.0 will provide superior image quality across all resolutions in all games it will support with quite impressive performance gains. However, FSR 2.0 would not need AI or utilize any sort of machine learning in order to improve the performance even more.
No AI or Machine Learning
DLSS works so well because on top of the temporal upscaling component, it also has a machine learning component that uses AI to aid in the boosting of frames. This is why DLSS is bound by NVIDIA’s proprietary hardware, DLSS needs the Tensor Cores present inside GeForce RTX graphic cards to gather and send data back to NVIDIA servers where the neural network is trained each time DLSS is used which, in turn, helps DLSS improve as a whole.
FSR 2.0 reportedly does not have an AI component at all. This can be interpreted as both a good and bad thing. Bad in the sense that it somewhat limits the maximum potential of the technology as AI can improve it even more. Good in the sense that because it does not need AI, it also does not require any proprietary hardware which allows it to work on a wide variety of hardware, including that from competitors.
While AMD does not confirm that no dedicated machine learning or AI-acceleration hardware is required for FSR 2.0, it does not clarify which GPUs the technology will actually work on. AMD only says that FSR 2.0 will support “a wide range of products and platforms, both AMD and competitors” without further confirmation. Even the footnotes for this sentence do not specify anything about competitor or AMD’s own hardware.
Lastly, Videocardz provided a picture of a cropped comparison slide that just shows the FPS, comparing native versus FSR 2.0 upscaled performance. As you can see below, the numbers are impressive, exhibiting a 2x improvement in FPS. However, do keep in mind that this slide shows the upscaled FPS for “Performance Mode” which is the mode that provides the most amount of FPS at the cost of image quality. Balanced or Quality modes should be better indicative of what FSR can do with less of an affect on the visuals.
So, we know that AMD FSR 2.0 will provide better performance than native, use temporal upscaling instead of spatial, won’t require AI or ML, and will work with AMD’s own GPUs along with a few from rivals. That being said, we are unsure if FSR 2.0 will be open-source like FSR 1.0 was and whether its code will be publicly available on GPUOpen. We’re also not clear what kind of hardware FSR 2.0 will run on, although RDNA 1 and 2 GPUs are a given.
This approach to upscaling games is rather interesting. AMD is going in a sort of hybrid direction where they are trying to give you the best of both worlds: temporal upscaling and no need for proprietary AI-acceleration hardware. Intel is trying to do something similar with XeSS as that will also work on competitor GPUs, is based on temporal upscaling, and won’t require AI, though it will benefit from it as XeSS will work best with dedicated hardware found on its own Arc Alchemist GPUs.
Some may even argue that by going in this direction, AMD is actually falling right in between instead of tipping on one side. Either support legacy hardware with easy open-source implementation or go for a more performant but limited approach that will produce more FPS, which is ultimately the fundamental goal of FSR, or any upscaling tech for that matter. Now, all we have to do is wait for the official announcement on March 17th and see how FSR 2.0 stacks up against the competition in real life.