Unofficial mod puts Nvidia's DLSS 5 neural rendering on Radeon, now back to RX 6000

Unofficial mod puts Nvidia's DLSS 5 neural rendering on Radeon, now back to RX 6000

A third-party project called DLSS-NR on AMD runs Nvidia's DLSS 5 neural rendering on Radeon graphics cards, and its October 3 release extended that to the RX 6000 series. It is unofficial software, labelled Alpha by the developer who publishes it, and neither Nvidia nor AMD has endorsed or supported it.

A library illustration of glowing letters falling towards a circular mirror. The project is unofficial, a user has to supply Nvidia's neural rendering library themselves, and neither Nvidia nor AMD has endorsed it. Image: Mixed News

The project is published on GitHub by a developer who goes by danielblnc, and its release history is unusually dense: nine builds between September 24 and October 3, each carrying a performance claim of its own. Its first release went out on September 3, and the repository held 2,300 stars when read at 00:25 UTC on October 5.

Version 0.6.0, published at 03:17 UTC on October 3, is the one that adds RDNA2 cards. Its notes say RX 6000 owners also need AMD's HIP 7.2 runtime installed separately, and list fidelity work alongside fixes for dark areas and small lights flickering in pre-upscale mode.

Every performance number here belongs to the developer

Nobody outside the project has published measurements of it, so each figure below is its author's claim about his own software. Version 0.5.0, on September 29, claims "up to 207% performance improvement over v0.4.3" on RX 7000 cards and 5% on RX 9000. Version 0.5.1, on October 1, claims a further 8% in Fast mode and 6% in Reference mode on RX 9000, and 3% in Fast mode on RX 7000.

The build before those, 0.4.3 on September 28, is titled "The 75 FPS update" and says "Performance now hits 75 FPS in Cyberpunk @ 1080p, 9070 XT Fast mode". That is one game at one resolution on one card in one mode, and it is the only frame rate the project quotes.

The two modes arrived in 0.4.2 on September 27. Fast "uses cheaper arithmetic where the difference is barely visible", in the developer's words, while Reference "keeps NVIDIA's exact arithmetic". RX 7000 cards always run Reference.

The pace held before that. Version 0.4.0 on September 25 claims 42% over 0.3.3, which had gone out 90 minutes earlier the same day claiming 23% over the build from the night before, and 0.4.1 on September 26 claims 8% plus a further 9% under heavy load from a new priority queue mode.

The project does not claim to match Nvidia's own implementation. Its summary says output quality is "very similar to what DLSS produces on the original hardware, though there are still some effects that may not be working 100%", and that performance remains a work in progress.

What it needs, and what it is not

The project targets DirectX 12 games that use FSR, on Windows 11, with Adrenalin driver 26.1.1 or later. It does not ship Nvidia's code: a user has to supply their own copy of the DLSS neural rendering library, build 310.8.0.0, from a game that already includes it. The project also warns that games with anti-cheat will block the file.

The weight of the thing has grown with the feature list. The installer went from 13.1MB at version 0.4.1 to 59.8MB at 0.6.0, and the newest build had been downloaded 7,020 times at that same read. GitHub does not flag any of these as prereleases, though the developer names every one of them Alpha.

Modders bending Nvidia's upscaling stack to places it was never shipped for is familiar ground here. Luke Ross brought DLSS Ray Reconstruction into four AAA games in stereoscopic VR in 2025, though that work stayed on GeForce cards.

What is different here is the vendor boundary. Neural rendering is the feature Nvidia sells with GeForce, AMD answers it with FSR, and a single developer is moving one onto the other's silicon in public, one release at a time. Neither company has said anything about it.

Leave a Comment

Your email address will not be published. Required fields are marked *