AMD on Monday, 28 October unveiled a real-time deep learning technique to produce high-quality graphics in its GPUs. Neural Supersampling introduced by AMD is said to challenge the widely used upscaling technology, DLSS of Nvidia.
AMD introduced supersampling along with a denoising technique to improve real-time path tracing. This might challenge Nvidia’s upper hand in AI-driven graphics technology. AMD in a blog post describes how neural supersampling and denoising work together to produce cutting-edge real-time path tracing.
The new technique involves upscaling and denoising with a single AI model. Nvidia’s DLSS technology has become a standard as it did both upscaling and denoising techniques that boosted performance in complicated ray-traced game scenes, whereas, AMD’s previous FSR technology, namely the latest FSR 3.1 has improved stability than the earlier models, but not the best in denoising path-traced frames.
Path tracing requires complicated computing as it uses raytracing to calculate all lighting. FSR 3.1 handles motion very well and delivers great picture quality. However, the noise created through incomplete illumination has to be cleared. FSR uses compute-based, hardware-agnostic solutions whereas DLSS uses highly improved AI-based iteration, and the latter falls in the first place.
AMD is following along with Nvidia and the latest model is claimed to challenge Nvidia’s DLSS. The company says, “Creating realistic images has been a persistently challenging problem in computer graphics, especially when it comes to rendering scenes with complex lighting.”
The post by the company states that they are developing and researching real-time path tracing on RDNA GPUs using AI neural networks. The major goals and research emphasis of the company are also detailed,AMD is investigating methods to “reconstruct high-quality pixels with fine details from extremely noisy images rendered at just one sample per pixel” and aims to “utilize minimal input by processing a single noisy color image rather than separate noisy diffuse and specular signals.” Additionally, AMD is exploring ways to “manage various types of noise from all lighting effects using a single denoiser, rather than relying on multiple denoisers for different effects.”
AMD aims for wide usability to support both denoising-only and denoising/upscaling modes from a single neural network and a 4K resolution which is highly optimized in performance for real-time path tracing.
AMD is focused on developing a Neural Supersampling and Denoising technique designed to produce high-quality, denoised, and supersampled images at display resolutions exceeding the original render resolution, all achieved through a single neural network for real-time path tracing.
The blog post that AMD posted on Monday explains the math behind the rendering and Monte Carlo integration, which is a modern technique used to ‘estimate the integrals with random samples in the domain’ and path tracing technology uses this technique to estimate the integrals in the equation. Path tracing is “unbiased and offers complex physically based rendering effects like reflections, shadows for a variety of scenes.”
Denoising is the process of improving the image quality by removing unwanted noise or grains from an image. AMD is rectifying the issue of noises with this new technology and clearly states to focus on it, hence the neural denoising technique. Neural denoisers are far better in technology compared to hand-crafted analytical denoising filters.
The neural network trained by AMD uses a large number of path-tracing images that can predict multiple filtering weights, decide on temporal accumulation, denoise, and upscale extremely noisy low-resolution images.
The new technique is able to denoise all noise and unwanted grains in a single pass as well as at low resolution and doesn’t need multiple denoisers which are used for different lighting effects in the rendering engine. The blog also shares images of the workflow of the AMD neural supersampling and denoising technique.
See more: AMD Struggles as NVIDIA’s RTX 3060 Leads GPU Rankings
