Upgrading technology from the “green” camp has long been strictly limited to GeForce cards, but the modding community has once again created a minor miracle. Developer nickname d4r proposed pilot projectwhich allows you to run official NVIDIA DLSS library versions 3, 4 or even 4.5 on AMD Radeon graphics accelerators.
The modification is based on the ZLUDA translation layer, which converts CUDA calls to execution via ROCm and HIP. The most interesting thing is behind the scenes – the heavy layers of the DLSS neural network were manually rewritten for the WMMA instructions and native FP8 computation in the RDNA architecture. The modifications are now focused on Linux and Proton environments, with the Navi 32 chip showing the best results, where custom shaders work many times faster than the base simulation. The game treats the graphics card as a DLSS-enabled solution, leaving processing to the original nvngx_dlss.dll library.
Of course, the project is still purely experimental – there are some flaws, and the setup process requires patience. Still, the precedent is impressive. After all, NVIDIA’s proprietary algorithms can now actually render frames on competing chips.
By the way, frame generation and DLSS 5 in this modification will not work, but as far as we know, DLSS 5 neural filters can already be launched on “red” graphics cards through other solutions.
