Unofficial DLSS 5 Mods Reach RTX 20/30 GPUs and PCSX2

Modders are expanding an unofficial, leaked build of NVIDIA DLSS 5 beyond its announced RTX 50 Series scope, with reported experiments on GeForce RTX 20 and RTX 30 GPUs, older graphics APIs, games without native DLSS, and the PCSX2 emulator. The work remains experimental and is not supported by NVIDIA.

Wccftech reports that the activity follows the appearance of an unfinished DLSS 5 Neural Rendering runtime through NBA 2K27. NVIDIA’s own DLSS 5 announcement presents the technology as a GeForce RTX 50 Series feature scheduled to launch officially this fall.

Older RTX cards use a different approach

Early attempts reportedly forced an FP8 workload onto older hardware with severe results. One cited Deep Rock Galactic test on an RTX 3080 fell from about 138 frames per second to 4 FPS after the leaked model was enabled.

A later community build is said to use FP16 on Turing and Ampere GPUs, making it a better match for RTX 20 and RTX 30 hardware. However, standardized benchmarks have not yet been published, so the scale of any improvement is unclear. Compatibility does not mean that older cards will deliver acceptable performance.

DLSS5-Feeder broadens game compatibility

The open-source DLSS5-Feeder project is designed to supply the depth and motion-vector data expected by the neural-rendering add-on when a game does not include native DLSS calls. Its documentation lists working paths for 64-bit Direct3D 11 and 12 games, a helper process for 32-bit titles, and Direct3D 9 through a wrapper. Recent project releases also add a Vulkan interoperability path.

Reported tests include Metro 2033 Redux, Splinter Cell: Blacklist, BioShock Remastered, and Fable Anniversary. A separate demonstration shown in the report applies the leaked model to Manhunt running through the PlayStation 2 emulator PCSX2.

This is not official DLSS 5 support

The distinction is important: NVIDIA has not announced DLSS 5 support for RTX 20 or RTX 30 cards, those older games, or PCSX2. The unofficial setup can carry a heavy performance cost, process interface elements along with the game world, produce motion-related artifacts, and behave differently from title to title.

The community work shows that the model can be adapted in unexpected ways, but it should not be treated as a preview of NVIDIA’s final compatibility list or performance targets. Until official software and independent benchmarks arrive, these experiments are best understood as technical demonstrations.