Nvidia overwhelmingly dominates the list of the best graphics cards, and that largely comes down to its feature set that’s been enabled through DLSS. AMD isn’t sitting idly by, however. The company is researching new ways to leverage neural networks to enable real-time path tracing on AMD graphics cards — something that, up to this point, has only really been possible on Nvidia GPUs.
AMD addressed the research in a blog post on GPUOpen, saying that the goal is “moving towards real-time path tracing on RDNA GPUs.” Nvidia already uses AI accelerators on RTX graphics cards to upscale an image via DLSS, but AMD is focused on a slightly different angle of performance gains — denoising.
When enabling path tracing in a game like Alan Wake 2 or Cyberpunk 2077, you’re only getting a small fraction of the rays cast into the scene. In a real-time context, only a handful of samples per pixel are cast into the scene, and they bounce around, but rarely go back to a light source within the scene. That leads to a noisy image — see the top left of the image above — that needs to be cleaned up with denoising. AMD is applying a neural network to the denoising process.
Nvidia has this technique covered already with Ray Reconstruction, which is a DLSS feature that’s sorely underrated. It makes a massive difference in image quality, preserving details in path tracing that would normally take minutes or hours to render for a single frame offline. AMD is looking at something similar: taking a small number of samples per pixel and reconstructing the fine details of path tracing using a neural network.
Get your weekly teardown of the tech behind PC gaming
The technique AMD is researching, however, combines upscaling and denoising into a single neural network. “We research a Neural Supersampling and Denoising technique which generates high-quality denoised and supersampled images at higher display resolution than render resolution for real-time path tracing with a single neural network,” the blog post reads. “Our technique can replace multiple denoisers used for different lighting effects in rendering engine by denoising all noise in a single pass, as well as at low resolution.”
This looks like some foundational research for the next version of AMD’s FSR, which could finally match Nvidia on performance and image quality. The lingering question is if these techniques require any bespoke hardware. Nvidia claims that dedicated accelerators on its RTX graphics cards are necessary for AI-assisted upscaling and denoising with DLSS, so AMD may need dedicated hardware on its GPUs, too.
However, there is a world where AMD could open up FSR 4 — or whatever the next version is called — to all graphics cards while still leveraging a neural network. RTX GPUs already have the hardware, and we’ve seen with features like Intel’s XeSS that it’s possible to run AI models on GPUs through separate instructions, though usually with a hit to image quality and performance.