The open-source project d4r has released an update enabling experimental support for NVIDIA DLSS 4 and 4.5 on AMD Radeon RX 9000 series graphics cards. This development allows AI super-resolution models designed for GeForce RTX hardware to run on the RDNA 4 architecture.
The integration relies on the RDNA 4 architecture's native support for FP8 calculations, which is the data format utilized by newer DLSS models. By processing FP8 directly, the project avoids the need to convert data to FP16, potentially reducing performance overhead and increasing efficiency.
The d4r project functions by utilizing the nvngx_dlss.dll file within Linux and Proton environments. It employs the ZLUDA compatibility layer to translate CUDA workloads into instructions compatible with AMD GPUs, supplemented by custom kernels developed for RDNA 3 and RDNA 4 architectures.
Currently, support for RDNA 4 remains in an early experimental phase. While versions 0.1.2 and 0.1.3 list cards like the RX 9070 XT and RX 9070 as targets, performance data is not yet fully verified. Testing has primarily focused on the previous RDNA 3-based Radeon RX 7700 XT.
