Nvidia's Innovations in Ray Tracing Technology

Method for accurately lighting virtual scenes in real-time rendering applications with complex, dynamic scenes. The method involves adaptive importance sampling of light sources in a scene using a cumulative distribution function (CDF). The CDF is generated based on importance estimates of lights in each voxel of the scene. This allows efficient sampling of lights for scene rendering by using the CDF to determine which lights are more important in a region. The CDF is updated as lights move or occlude to adapt to changing lighting conditions. Jittering voxel boundaries and light normal directions helps smooth out artifacts.

Source

Accelerating ray tracing and material shading on parallel processors like GPUs without adding transistors. The method involves compiling material graphs into optimized byte code instructions for execution on the parallel processing cores. The material graphs represent the surface properties and relationships of objects. The byte code is generated by parsing the graphs into expression trees and optimizing them. This allows efficient evaluation of the material graphs using the parallel processing cores without branching or sorting rays.

Source

Spectral rendering of images with volume media like mist, fog, dust, clouds, water, etc., that closely matches rendering in a multi-color scheme like RGB. The method involves identifying a light in a scene, determining an attenuation function characterizing interaction with the medium, fitting parameters based on minimizing color difference, and rendering using the attenuation function and fitted parameters. This allows realistic spectral rendering of lights transmitting through media that matches the appearance in a multi-color rendering.

Source

Render shadows in scenes with foliage like trees and bushes using a hybrid approach that improves performance and realism when ray-tracing shadows of alpha-tested geometry. The approach involves combining noisy visibility samples from raytraced opaque geometry with shadow maps that only contain alpha-tested geometry. This results in hybrid denoised shadows that show penumbras of high quality and hard ray-traced shadows at contact points with a low performance impact. The alpha-only shadow map only contains depth values from geometry that passes the alpha-test.

Source

Optimizing motion vector estimation for translucent objects in images to improve rendering quality and reduce noise. The optimization involves accurately finding motion vectors for pixels in a current frame by searching for matching world positions in previous frames, even when objects are transparent. It uses numerical optimization techniques like Newton's method to minimize the angle between vectors instead of finding closest pixels. This improves performance over naive methods like random walks. The optimized motion vectors are then used for tasks like denoising and temporal anti-aliasing.

Source

Generating realistic water caustics in real-time gaming and animation applications that appear and move like actual physical caustics, while avoiding the need for expensive denoising procedures that can be computationally expensive or prohibitive for real-time use. The method involves using a low resolution caustics map with each point treated as a vector. These vectors are rendered as a mesh with connectors between adjacent points. When a hit point is detected, the position of the hit point replaces the original vertex position. This projected mesh provides sharp caustic patterns using fewer points compared to traditional denoising. The density of triangles is determined based on area to accurately represent brightness.

Source

Rendering signed distance functions (SDFs) in computer graphics more efficiently by improving techniques for tracing rays, computing surface normals, and finding ray intersections with SDF surfaces. The techniques involve: 1. Factorizing SDF coefficients into parameters to compute ray intersections with SDF surfaces using a cubic function. 2. Computing surface normals by interpolating neighboring voxel normals. 3. Finding ray intersections with SDF surfaces using turning points of the cubic function to determine shadowed regions. These techniques improve ray tracing and surface computation efficiency compared to duplicative calculations and separate ray tracing for shadows.

Source

Interleaving textures to improve ray tracing performance for scenes with incoherent light rays. Instead of separately allocating and accessing cache lines for each texture associated with an object intersected by a ray, the textures are combined into a single interleaved texture. This reduces the number of cache lines needed and less texture data is read compared to separately accessing each texture. The interleaved texture contains blocks from multiple textures alternating in order. It can be accessed using multiple texture headers with stride distances between blocks from the same texture. This allows fetching from the interleaved texture instead of separate textures for rays with incoherent intersections.

Source

Hardware acceleration of ray tracing for realistic graphics rendering that provides selective execution of programmable ray operations. The technique allows efficient and flexible ray intersection tests by enabling custom ray operations like selecting the level of detail (LOD) of an object. This improves ray tracing performance by allowing optimized ray operations for acceleration structures like BVHs that have mixed LOD models. It avoids the rigidity of requiring all nodes to support the same programmable ray operations. Instead, the ray operations can be applied selectively to nodes based on their data.

Source

Detecting penumbra regions for shadow denoising using wave intrinsic functions on parallel processors to avoid expensive post-processing for penumbra detection. The approach leverages threads of schedulable units (e.g., warps) used for visibility sampling during ray tracing to identify penumbra regions. Each thread computes a value indicating if its pixel region is in a penumbra using intrinsic functions. This avoids post-processing and allows selective denoising of penumbra regions vs fully lit/shaded regions. It also determines filter parameters based on sampling statistics to adapt denoising for regions.

Source

Hardware acceleration for real-time ray tracing to enable interactive ray tracing for applications like virtual reality and augmented reality. The acceleration is provided by a dedicated co-processor called a traversal coprocessor that sits in the graphics pipeline alongside the streaming processors. The coprocessor efficiently traverses an acceleration data structure like a bounding volume hierarchy to quickly determine ray intersections with scene objects. It handles tasks like ray-bounding volume intersection, ray-primitive intersection, and ray transforms. The coprocessor's dedicated hardware provides faster ray tracing compared to software ray tracing.

Source

Hardware-accelerated ray tracing that provides flexible and efficient ray intersection tests for real-time graphics. The technique allows combining ray operations like instance masking and geometric level of detail testing in parallel processing units. This improves ray tracing efficiency by enabling more complex ray tracing effects while simultaneously improving ray tracing efficiency. It allows dynamically choosing ray traversal based on multiple selection criteria per ray. The ray operations are configured to occur before and after node stack push/pop in the traversal coprocessor. This enables combining ray operations like instance masking and geometric level of detail testing in parallel processing units.

Source

Efficiently generating high quality images using techniques that allow efficient sampling of large emissive textures for lighting scenes. The method involves determining cumulative distribution functions for textures, constructing geometric representations from them, and tracing rays against the geometry to sample the textures efficiently. This hardware-accelerated sampling reduces noise compared to naive random sampling. The approach allows realistic lighting of scenes using large textures without incurring excessive computation cost.

Source

Rendering techniques for media like surfaces with microscopic irregularities and homogeneous slabs using position-free path integrals. It involves sampling directions of light passing through the medium, computing parameters for position distributions based on the directions, computing brightness at the exit direction using the distribution parameters, and rendering images based on the brightness. This approximates path integrals through the medium without the noise of Monte Carlo simulations.

Source

Techniques for sharing irradiance between ray interactions spatially and temporally to improve ray tracing performance for rendering virtual scenes. The method involves using irradiance caches to aggregate and interpolate irradiance samples instead of sampling from every location. Irradiance caches are associated with locations and updated by casting fewer rays based on ranking factors. Irradiance from other caches can be blended to compute irradiance at a location. This reduces the number of samples needed for lighting calculations compared to full ray tracing.

Source

High-performance data communication system for scaling memory bandwidth in multi-GPU and multi-CPU systems. The system uses hardware page faulting to enable transparent copying between processors without pinning memory. This allows flexible transfer of data between GPUs and CPUs without reducing available memory. The system uses a memory partitioning unit that services page faults for unresident addresses, allowing copy engines to transfer data between processors without worrying about memory residency. This allows efficient scaling of memory bandwidth in high-performance systems with multiple processors.

Source

Reducing temporal lag in real-time ray tracing and other dynamic scenes without compromising image quality. It uses a fast history buffer with higher blend weight and clamping to determine a clamping window for the full history buffer. This allows controlling the balance between noise reduction and temporal lag. The fast history buffer is used to clamp the full history value before reprojection to the current frame. The full history is still maintained for accurate smoothing in static scenes without clamping. This fast-full history clamping technique reduces temporal lag in dynamic scenes without adding computational complexity compared to conventional methods.

Source

Improving surround view systems for vehicles by reducing artifacts, better representing useful visual information, and optimizing streaming of environment visualization. Techniques include: 1. Dynamic seam placement in stitched images based on object saliency and ego-object state to avoid important objects and reduce distortions. 2. Adaptive 3D bowl modeling of the surrounding environment based on distance to objects, providing better representation and avoiding distortions. 3. Reconstructing the area under the vehicle using cached sensor data and ego-motion to provide a complete visualization. 4. Streaming optimized visualization of the environment based on ego-object speed and direction to reduce data requirements.

Source

Avoiding self-intersections when rendering scenes with signed distance functions (SDFs) by compactly storing hit points in a grid representation. Instead of storing the full hit point coordinates, only an identifier of the grid voxel and subsets of the coordinates are stored. This reduces memory requirements and allows more accurate reconstruction of hit points for continuation rays. The missing coordinate values are computed based on the known ones and the voxel's canonical frame.

Source

Noise-free differentiable ray casting for optimizing 3D models using reference images. The technique involves casting multiple rays from each 3D point to compute visibility gradients. These gradients are backpropagated to update model parameters and reduce differences between rendered model and reference images. An analytical function renders the model with lower noise compared to stochastic sampling. This allows accurate 3D model reconstruction from reference images without requiring large ray counts.

Source