Experimental d4r Brings DLSS 4.5 to Radeon RX 9000 GPUs on Linux

The experimental open-source d4r project has expanded its Radeon support to AMD's RDNA 4 architecture, adding build targets for Radeon RX 9000-series GPUs. d4r runs NVIDIA's official DLSS Super Resolution library in Windows games under Linux and Proton, translating CUDA work through ZLUDA while replacing some of the heaviest neural-network kernels with hand-written Radeon code.

The current project supports the DLSS 3 CNN model, DLSS 4 transformer model and DLSS 4.5 transformer model. This is an unofficial compatibility experiment rather than AMD or NVIDIA support for running DLSS on Radeon hardware.

The technical change is RDNA 4's native FP8 matrix path. d4r marks NativeFp8 as the default for its gfx1200 and gfx1201 targets, whereas its RDNA 3 path widens FP8 work to FP16. That removes one translation step for the newer transformer models and gives the project a more direct way to execute their matrix operations on RDNA 4.

That architectural advantage is not yet a measured performance advantage. The developer explicitly says RDNA 4 runtime and performance remain unverified on physical hardware, so there is currently no defensible RX 9070 XT or RX 9060 XT frame-rate claim.

On the developer's Radeon RX 7700 XT test system in Silent Hill: Townfall at 2560×1440, d4r reports 49.3 FPS for DLSS 4.5 Quality, 67.9 FPS for DLSS 4 Quality and 71.7 FPS for DLSS 3 Quality. The project's earlier-session FSR 4 result was 76.0 FPS, while native-resolution TSR at 100% was 49.1 FPS.

Those numbers are developer measurements rather than independent Core Tech Tips benchmarks, and the project notes that the FSR 4 run came from an earlier session when the machine was around 2–4% faster overall. They are useful for showing the present experimental cost, not for declaring a clean cross-upscaler winner.

d4r currently targets DLSS Super Resolution. It does not turn a Radeon GPU into a fully supported GeForce DLSS platform, and the project does not provide DLSS Frame Generation, Ray Reconstruction or DLSS 5 Neural Rendering.

The software path is also specific: Windows games run under Linux and Proton, with ZLUDA, ROCm, patched vkd3d-proton and project-specific native kernels involved. The developer describes d4r as a proof of concept that is not yet practical for everyday use and warns that some game/model combinations can show visible artifacting.

The useful result is therefore narrower than 'DLSS now works on Radeon.' d4r demonstrates that NVIDIA's recent Super Resolution networks can be made to execute on AMD hardware through an experimental compatibility stack, while the new RDNA 4 code path shows how native FP8 hardware could reduce part of that translation burden once real RX 9000 testing exists.