Thursday, March 18th 2021
Confronting NVIDIA's DLSS: AMD Confirms FidelityFX Super Resolution (FSR) to Launch in 2021
AMD, via its CVP & GM at AMD Radeon Scott Herkelman, confirmed in video with PCWorld that the company's counterpart to NVIDIA's DLSS technology - which he defines as the most important piece of software currently in development from a graphics perspective - is coming along nicely. Launch of the technology is currently planned for later this year. Scott Herkelman further confirmed that there is still a lot of work to do on the technology before it's ready for prime time, but in the meantime, it has an official acronym: FSR (FidelityFX Super Resolution). If you're unfamiliar with DLSS, it's essentially an NVIDIA-locked, proprietary upscaling algorithm that has been implemented in a number of games now, which leverages Machine Learning hardware capabilities (tensor cores) to upscale a game with minimal impact to visual quality. It's important because it allows for much higher performance in even the latest, most demanding titles - especially when they implement raytracing.
As has been the case with AMD, its standing on upscaling technologies defends a multiplatform, compatible approach that only demands implementation of open standards to run in users' systems. The idea is to achieve the broadest possible spectrum of game developers and gamers, with tight, seamless integration with the usual game development workflow. This is done mostly via taking advantage of Microsoft's DirectML implementation that's baked straight into DX 12.
One detail doesn't instill confidence in how soon we'll see this technology out in the wild; Scott Herkelman in the video says that there are multiple approaches to such an upscaling solution, and that they're being evaluated in the lab; this either means that AMD hasn't yet decided on the technologies to leverage for the upscale via Microsoft's Direct ML, or that the company is actively working on two or more different approaches to actually be able to measure their benefits, drawbacks, and ability for deployment in a large scale. All in all though, it's great to know that things are coming along nicely, as such a technology has an immense return potential not only for PC gamers (perhaps even NVIDIA-toting ones, if AMD's solution truly is hardware agnostic), but also for console players. If the performance increases we can expect from FSR are comparable to those of DLSS, we can expect an immense amount of power being unlocked in current-gen consoles. And that, in turn, benefits everyone.
Watch the full PCWorld video below:
Source:
via Videocardz
As has been the case with AMD, its standing on upscaling technologies defends a multiplatform, compatible approach that only demands implementation of open standards to run in users' systems. The idea is to achieve the broadest possible spectrum of game developers and gamers, with tight, seamless integration with the usual game development workflow. This is done mostly via taking advantage of Microsoft's DirectML implementation that's baked straight into DX 12.
Watch the full PCWorld video below:
89 Comments on Confronting NVIDIA's DLSS: AMD Confirms FidelityFX Super Resolution (FSR) to Launch in 2021
also dont forget that the consoles are RDNA2 so devs will most def work to make the tech work for those and therebye easily can port that over to the PC version
We'll have to wait and see how the situation with DLSS and FSR develops.
Here's a link for those wanting to read from the source.
www.amd.com/en/technologies/radeon-software-fidelityfx
People have been trying to debug or analyze GPU usage when DLSS is used since the beginning. It is not just running on shaders. Tensor cores are used, whether they are doing matrix/ML stuff or FP16 shaders that they are reportedly also used for, is unknown.Not sure if DLSS or FCR is at the same level as API.
Anyway, marketing BS goes with it regardless of whether the solution is standard or not. See, it is called FCR.
What AMD has been struggling with for a while is to come up with new usable technology or effect and get it in widespread enough adoption first. When you are second and somebody is already offering and marketing something similar, proprietary or not, it is an uphill battle. Mantle/DX12/Vulkan was probably the last limited success in that regard. Also Async shaders but the actual gain/effect from that is pretty limited - I think VRS is a good comparison to that with a similar result. Vega brought cool stuff but it largely remained unfinished, perhaps with the exception of RPM which has limited adoption even with bipartisan support by now.
SAM might be an interesting example - remember that it was supposed to be limited to Ryzen 5000 series on 500-series boards and Radeon 6000 series? After Intel made efforts to enable this on some of their platforms and Nvidia announced it will enable something similar this is set to become a very general thing.
I disagree, I think higher refresh is welcome on all platforms, but time will tell.
Open source usually means a source code that is made freely available for possible modification and redistribution. Nothing about how easily or hard is it is to add it to game core engine while game engine plugin usually means you only have to enable it inside the game engine. Good luck implementing DirectML API or some other type of accelerating machine learning if all rendered things on screen don't have motion vectors tied to them at all distances... You have to fix that first (A LOT OF WORK), while plugin inside base game engine "just works" (well not exactly, but we're talking about losing weeks not months to make it work right), because it's optimized for it and vice versa. We need super simple plugin solutions for all main game engines, be it from Microsoft, Intel, Nvidia or AMD in order to get them into the PC games.
The actual APIs can be as open as possible, but there's still a learning part involved in the training phase. In a completely open solution who performs the training? Who acts as a repository for the training results?
Sure, chances are we'll end up with an open API after all (even DXR was open from the beginning, even if Nvidia tacked on RTX). I'm just saying, the meat of these algorithms will remain in AMD's and Nvidia's hands.
And don't even get me started on bs 4K DLSS enabled Switch rumors.
Cyberpunk 2077 is an Nvidia sponsored title and it includes CAS with dynamic and static resolution from the FidelityFX suite.RDNA2 does support the acceleration of AI anyways, they just don't need separate hardware to do it.
You are also missing the point that AMD pretty much cornered the console market (less Nintendo) at this point. A good upscale algo that is easy to implement is great for game makers to utilize for consoles, because instead of trying very hard to squeeze performance out by means of optimizing the codes, they can reinvest that time and utilize upscaling to gain performance. For example, if you look at Cyberpunk on older hardware, if you can upscale the game from 720p to 1080p with no severe degrade in quality, it will make the game from unplayable to playable. Not sure if this works on older AMD hardware, but just quoting an example.Technique should be the same, but tweaks required depending on the OS. XBox is pretty much running Windows Lite, so should not have a lot of rework. But if we look at PS5, then I suspect there should be quite a fair bit of rework in terms of coding. I don't think they will want to leave Sony out in the cold.
Secondly although PS5 and XBSX use semi custom RDNA2 SOC, there are some design differences between PS5, XBSX, RX5000 and RX6000 regarding low-precision operations that are used for AI image enhancing. Redgamingtech provide some information here
With XBSX SOC share the most common with RX6000 GPU, yeah sure maybe XB ported games will include FSR that is optimized only for RX6000 GPU. PS ported games however would probably go DLSS, as many of them already did such as Death Stranding, Monster Hunter World, FF XV, Nioh 2, etc....For independent PC developers, I don't see any incentive for them to ignore DLSS which has a much bigger user base and go for FSR which has almost non-existant user base atm.
Cyberpunk 2077? I wouldn't play that game without at least an RTX2060, which trashes the 1080Ti when using DLSS (and DLSS looks miles ahead of the FidelityFX option)
this game can run on a 1060 if you TURN THE GRAPHICS DOWN