A GitHub project loads NVIDIA’s neural renderer onto RDNA 4 and RDNA 3 silicon. It works. It is also extremely slow.
Three days after DLSS 5 shipped in its first retail game, someone has it running on a Radeon. DLSS 5 on Radeon is not a spoof, not FSR wearing a costume, and not a rumour: it is NVIDIA’s own neural rendering DLL executing on AMD hardware, and you can download it right now.
The project is DLSS-NR-on-AMD, published by a modder called Daniel Blanco. It targets Radeon RX 9000 (RDNA 4) and RX 7000 (RDNA 3) cards in any DirectX 12 game that already exposes FSR, and it has collected more than 700 GitHub stars in a couple of days.
How it gets there #
The clever part is that the mod never lets NVIDIA’s runtime ask what card it is talking to. It sits in front of the NGX API, intercepts the calls and loads the neural rendering DLL regardless — then, instead of handing the work to CUDA, it translates the maths onto AMD’s HIP RT backend. The game’s existing FSR pipeline supplies the frames. As far as the game is concerned, you simply switched FSR on.
Setup is about as rough as you would expect from an alpha. You drop a setup executable into the game’s binary folder, supply your own copy of nvngx_dlssnr.dll build 310.8.0.0 pulled from a title that already ships DLSS 5, run the installer, then enable FSR in-game and hit End for an overlay with sliders for tone intensity, structure intensity and skin structure. Requirements are Windows 11 and Adrenalin 26.1.1 or newer. Blanco has tested it on Cyberpunk 2077 and GTA V Enhanced. Vulkan is on the list but not there yet.
Now the frame rate #
This is where the magic trick stops being useful. The README is refreshingly blunt about it: roughly 33 FPS at 1080p on an RX 9070 XT. Earlier builds were worse still, with the first reports having Cyberpunk 2077 crawl along at 11 to 12 FPS before successive alphas clawed their way into the high twenties. Blanco says the goal is to eventually match an RTX 5070 Ti, which is an enormous distance from where the code sits today.
Image quality, by contrast, is apparently close. The repository claims output very similar to what DLSS produces on the original hardware
, with the caveat that some effects still are not working properly. That is the genuinely interesting half of the result: the neural network itself runs fine on RDNA silicon. What is missing is the tensor throughput and the hand-tuned scheduling that make it cheap on a GeForce card.
It is worth remembering that DLSS 5 is expensive even on the hardware it was written for. When it shipped on September 3 in its first supported title, it was still costing around half the frame rate on native NVIDIA silicon. A port that lands at a fraction of that on a mid-range Radeon is not really a shock.
The caveats that matter #
Anti-cheat blocks the DLL outright, so anything with kernel-level protection is off the table unless you are playing offline with it disabled. You are supplying NVIDIA’s proprietary DLL yourself, which is its own grey area. And alpha here means alpha: RX 7000 support is listed as untested, and Blanco is openly asking for reports.
None of which makes the result less impressive. Vendor-locked upscalers have been the norm for years, and the working assumption has been that DLSS stays on GeForce because it has to. This says otherwise. The lock is a check, not a wall. Whether anyone actually wants to play like this is a separate question, but the proof of concept is now sitting in a public repository, and the star count suggests plenty of people are curious.