AMD Develops Neural Supervisual Processing Techniques to Combat Noise in Ray Traced Frames

Company AMD is working to improve its image quality technologies by developing neural supervised visual processing techniques to remove noise from ray-traced frames. Such developments should improve visual effects, approaching the quality of the solution Ray Reconstruction from NVIDIA, which uses Tensor cores for more accurate noise reduction.

NVIDIA, thanks to DLSS and Ray Reconstruction, has significantly improved the performance and visual quality of modern games, which has also prompted AMD to develop similar solutions. The current technology FSR (FidelityFX Super Resolution) has been upgraded to version 3.1, which already solves many stability issues, but noise reduction problems remain. Now AMD intends to implement neural network methods on its GPUs of the series rDNA, which will allow for better noise removal and improved visual effects during rendering.

Unlike NVIDIA, which uses Tensor Cores, AMD relies on WMMA (Wave Matrix Multiply Accumulate), which should make their approach more versatile. AMD says it is actively researching neural noise cancellation for the technology. Monte Carlo, which should bring them closer to real time when ray tracing.

It is expected that these methods can replace several separate denoisers working on different lighting effects in a render, performing this task in a single pass. In the future FSR4.0 may have limited compatibility with video cards NVIDIA и Intel, but will offer full functionality only for rDNA. CTO AMD Mark Papermaster has already confirmed that they are implementing AI-powered upscaling technologies in their gaming devices, which indicates the company's long-term strategy in this area.