AMD could swipe some of the best features of Nvidia GPUs

Jacob Roach / Digital Trends

Nvidia overwhelmingly dominates nan database of nan best graphics cards, and that mostly comes down to its characteristic group that’s been enabled done DLSS. AMD isn’t sitting idly by, however. The institution is researching caller ways to leverage neural networks to alteration real-time path tracing connected AMD graphics cards — thing that, up to this point, has only really been imaginable connected Nvidia GPUs.

AMD addressed nan investigation successful a blog station connected GPUOpen, saying that nan extremity is “moving towards real-time way tracing connected RDNA GPUs.” Nvidia already uses AI accelerators connected RTX graphics cards to upscale an image via DLSS, but AMD is focused connected a somewhat different perspective of capacity gains — denoising.

AMD

When enabling way tracing successful a crippled like Alan Wake 2 or Cyberpunk 2077, you’re only getting a mini fraction of nan rays formed into nan scene. In a real-time context, only a fistful of samples per pixel are formed into nan scene, and they bounce around, but seldom spell backmost to a ray root wrong nan scene. That leads to a noisy image — spot nan apical near of nan image supra — that needs to beryllium cleaned up pinch denoising. AMD is applying a neural web to nan denoising process.

Nvidia has this method covered already pinch Ray Reconstruction, which is simply a DLSS characteristic that’s sorely underrated. It makes a monolithic quality successful image quality, preserving specifications successful way tracing that would usually return minutes aliases hours to render for a azygous framework offline. AMD is looking astatine thing similar: taking a mini number of samples per pixel and reconstructing nan good specifications of way tracing utilizing a neural network.

The method AMD is researching, however, combines upscaling and denoising into a azygous neural network. “We investigation a Neural Supersampling and Denoising method which generates high-quality denoised and supersampled images astatine higher show solution than render solution for real-time way tracing pinch a azygous neural network,” nan blog station reads. “Our method tin switch aggregate denoisers utilized for different lighting effects successful rendering motor by denoising each sound successful a azygous pass, arsenic good arsenic astatine debased resolution.”

This looks for illustration immoderate foundational investigation for nan adjacent type of AMD’s FSR, which could yet lucifer Nvidia connected capacity and image quality. The lingering mobility is if these techniques require immoderate bespoke hardware. Nvidia claims that dedicated accelerators connected its RTX graphics cards are basal for AI-assisted upscaling and denoising pinch DLSS, truthful AMD whitethorn request dedicated hardware connected its GPUs, too.

However, location is simply a world wherever AMD could unfastened up FSR 4 — aliases immoderate nan adjacent type is called — to each graphics cards while still leveraging a neural network. RTX GPUs already person nan hardware, and we’ve seen pinch features for illustration Intel’s XeSS that it’s imaginable to tally AI models connected GPUs done abstracted instructions, though usually pinch a deed to image value and performance.