Denoiser Engine, the Return of Dual GPU Graphics Cards

One of the problems of Path Tracing or better known as Monte Carlo Ray Tracing is the number of samples that are necessary per pixel to achieve a sharp image, that is, without noise of any kind, which requires a computational power outside beyond the reach of home computers.

In Path Tracing, the rays are distributed randomly in each pixel of the scene, when an intersection occurs then an indirect ray is generated (reflection, shadows, etc.) that points to a random direction. Thus, after a few bounces, the ray leaves the scene or is absorbed by striking an object with a refractive quotient of 0 or close to 0. When each of the rays has finished bouncing around the scene, then it is calculated a sample value, which is based on the information obtained from the beam’s path through the scene.

The fact that the distribution of the rays is done in a random way causes that a large number of samples are necessary, which causes that a huge calculation capacity is needed to have a clear image in a scene using Path Tracing, to the point where that CG movies using this technique require powerful supercomputers.

The Denoiser Engine

Such is the computational load of this type of rendering that companies like Disney and its subsidiary Pixar decided a few years ago to start using Deep Learning networks for the reconstruction of images so that they imagine / hallucinate the complete images of their films made in CG using a very few samples. Thus, with an image with much fewer samples per pixel, the same result is achieved as one with many samples per pixel using a “Denoising” or noise elimination algorithm through Deep Learning and dedicated processors applying this algorithm to the images. produced by a GPU with far fewer samples.

There is a Disney / Pixar patent entitled Multi-scale architecture of denoising monte carlo renderings using neural networks where precisely they talk about the Denoiser Engine as a concept and I think it is something that is going to be standardized in the world of 3D graphics, not face to games, but in terms of composition of 3D images.

That is, the idea is to assign a coprocessor to the GPU specialized in executing neural networks and prepared to carry out the Denoising process, said coprocessor in its first versions if we apply it to contemporary GPUs can be connected to the GPU through an interface NVLink (NVIDIA) or an xGMI (AMD) interface within the same graphics card and receive the noisey image buffer from a DMA mechanism to copy it to its memory and apply the noise elimination process on the image generated by the GPU.

The Denoiser Engine could mean a return to dual GPUs

Contemporary GPUs, at least those from NVIDIA, have a number of special drives to speed up Artificial Intelligence algorithms. In the case of AMD these units are in the CDNA architecture of their recently launched AMD Instinct MI100.

This can lead to the return of dual GPUs, but in which while one renders the scene for Path Tracing, the other GPU is responsible for applying the AI denoising algorithm. The idea is therefore very similar to that of DLSS, but instead of using the Tensor Cores to generate an image at a higher resolution, they are used to reduce the noise of the scene and to be able to achieve a clear image with fewer samples .

The usefulness of the Denoiser Engine in games

There is still a long time to go before we see the exclusive use of Ray Tracing in games, or failing that, we will see a greater number of games that are purely based on it and not on a combination of ray tracing and rasterization. But we have cases like the Minecraft RTX that are rendered through pure ray tracing.

As you can see, the process of rendering the Minecraft RTX ends up spending much more time removing noise from the scene than rendering it. As more games appear that use Ray Tracing to render the scene, it will be more necessary to perform the denoising process efficiently, hence the need for a Denoiser Engine in the medium-long term.