AMD Neural Supersampling Aims to Compete with Nvidia DLSS

AMD's research team has introduced a new step forward in real-time path tracing, launching a fresh neural network method that merges both denoising and upscaling. This initiative is directly aimed at competing with Nvidia's dominance in AI-based graphics technology.

Tackling Challenges in Path Tracing

The innovative technology addresses one of the main challenges faced in real-time path tracing: achieving high-quality images with only a few ray samples per pixel. Typically, path tracing requires thousands of rays per pixel to produce those intricate frames, explaining why rendering a single frame in films can take hours. AMD's approach emphasizes reconstructing detailed scenes from limited samples using neural network processing.

A Unified Process

The neural network developed by AMD effectively combines denoising and upscaling into a singular process. In most rendering systems, these steps are usually separate, but AMD has streamlined the workflow. By processing low-resolution renders with merely one sample per pixel and utilizing temporal accumulation and guidance buffers, their system can recreate sharp, high-resolution visuals.

Comparing with Nvidia's Technology

This technique is not quite the same as Nvidia's DLSS, which divides upscaling, frame generation, and ray reconstruction into distinct phases. DLSS also relies on specialized AI hardware found in RTX GPUs, but it remains uncertain whether AMD's solution will function on existing RDNA GPUs or if new hardware will be necessary for support.

AMD's team has outlined several key objectives they hope to achieve with this new technology:

This breakthrough may serve as the foundation for the upcoming next-gen FidelityFX Super Resolution (FSR) from AMD. Nevertheless, considering how demanding modern path-traced games are, we will probably require more powerful hardware to fully utilize the capabilities this technology offers.

Leave a Comment

Scroll to Top