Nvidia showcases neural beam caching technology to improve global illumination

During the SIGGRAPH 2021 virtual event, nvidia He presented some of his latest research on real-time graphics progression. Without a doubt, the most interesting show was the show Neuroradiation bufferingA new technology designed specifically for global illumination by tracing.

Neural Radiance Caching combines RTX (Nvidia Tensor Cores) Neural Network Accelerator hardware and RayTracing (Nvidia RT Cores) hardware to create a system capable of fully dynamic global illumination that works with all types of materials, whether opaque, glossy or volumetric. Accurate textures such as whiteness, roughness or terrain maps, and adapts to large outdoor environments without the need for auxiliary data structures or scene settings.

Combined with Nvidia’s next generation direct illumination algorithm, ReSTIR, neural radiation caching can improve global illumination display efficiency by up to 100 factors – two orders of magnitude.

At the heart of this technology is a single small neural network that works up to 9 times faster than TensorFlow v2.5.0. Its speed enables live training on the network during gameplay to keep up with arbitrary dynamic content. On the Nvidia GeForce RTX 3090 graphics card, caching can provide Neural Radiance More than 1 billion global lighting queries per second. “

Nvidia graphics researchers have also worked on what’s called a neural reflex field texture, or Nerf Tex In short, his goal More precisely shape complex materials such as leather, cloth or grass.

Instead of using classic graphic primitives to model the structure, we propose the use of a versatile volumetric primitive represented by the Neural Reflective Domain (NeRF-Tex), which co-designs the material’s geometry and its response to illumination. It can be instantiated on a core network to “configure” it to an intermediate appearance. ​and the required microscale. We adapt the reflection field to user-defined parameters that control appearance. Thus, a single NeRF texture captures an entire area of ​​reflection fields rather than a specific structure. This increases the range of appearances that can be modeled and provides a solution to combat artifacts of repetitive motifs.”

via: Wccftech