Graphics & Rendering

Arm received 2 patents focused on bringing frame interpolation to mobile hardware. One describes a neural network that learns optimal blending parameters to generate intermediate frames, producing smoother motion without requiring the GPU to render each frame from scratch. The second extends this into a full pipeline designed around the thermal and bandwidth constraints of smartphone and tablet processors, addressing a gap left by existing frame generation technologies that were built for high-power desktop and console GPUs.

Nvidia received 1 patent that rethinks how GPUs handle concurrent memory operations. Rather than relying on a fixed set of atomic operations baked into hardware, the approach allows programmers to define their own custom atomic operations in software, executed by specialized processors within the memory hierarchy. This removes the dependency on long hardware development cycles when new types of concurrent operations are needed.

Qualcomm's 1 patent addresses variable rate shading on mobile GPUs through a predictive rather than reactive mechanism. By analyzing draw call data from previous frames, the system anticipates which upcoming frames are likely to be GPU-heavy and reduces shading rates in advance, before a performance drop occurs. Existing variable rate shading implementations adjust after a bottleneck is already detected, making this lookahead approach a meaningful departure from the standard method.

Nintendo received 2 patents covering distinct areas of game rendering. One describes a voxel mesh system where terrain can be deformed or destroyed by player actions, with a dual-mode approach that preserves intersection data for reversible deformations while discarding it for permanent destruction, allowing high-fidelity mesh reconstruction without excessive memory use. The other patent addresses a common problem in racing games, where distant competitors become difficult to see, by enlarging those objects at the vertex shader level in proportion to their distance, keeping them visible without adding interface clutter or relying on post-processing.

Tencent's 1 patent accelerates compressed texture decoding by moving the process entirely onto the GPU. The system assigns one GPU work group to each individual compression block, allowing every block to be decoded in parallel simultaneously rather than sequentially. This removes the CPU from the decode pipeline entirely.

NetEase received 1 patent for a technique that creates the appearance of depth in character display backgrounds using stacked 2D texture layers rather than full 3D scene geometry. The sizes of each layer are calculated based on the viewing frustum so that every layer correctly fills its plane in clip space, producing a convincing parallax effect. This achieves the visual result of a three-dimensional background at a fraction of the computational cost.

Sony received 5 patents spanning several areas of rendering and graphics processing. One combines a conventional mesh pipeline with a neural network operating at the fragment shader level, where the mesh handles geometry and the network focuses only on surface appearance, reducing what the neural model must learn compared to a full neural radiance field approach. A second patent tackles performance stability during GPU-intensive scenes by dynamically switching between pre-cached versions of an AI upscaling model at different precision levels, maintaining frame rate without reducing output resolution. A third applies a similar distribution logic to spectral ray tracing, spreading wavelength samples across neighboring pixels and successive frames rather than computing multiple wavelengths per pixel per frame, reducing workload while preserving perceived visual quality. The fourth patent performs a lightweight partial render before streaming any texture data, using the result to identify exactly which textures and mipmap levels are needed for the current view, avoiding transfers of texture data that would never appear on screen. The fifth uses Gaussian splatting to construct a volumetric representation of a game scene from video and metadata, allowing user annotations such as coaching diagrams to be placed in three-dimensional space with correct depth ordering and occlusion rather than as flat 2D overlays.

Cyber Radiance received 1 patent for a real-time hair rendering approach that generates individual strands procedurally each frame from compact hair mesh structures, rather than retrieving pre-computed strand models from memory. Styling operations are applied within a local coordinate system that moves with mesh deformation, so the hair responds correctly to character movement. The approach is designed to be faster than memory retrieval of stored models while giving artists meaningful control over the final appearance.

Intel's 1 patent solves a specific problem that arises when displays switch between resolution and refresh rate modes, such as moving from 4K at 240Hz to 1080p at 480Hz. Normally this transition causes several seconds of screen blanking while the display resynchronizes, but the patent describes padding lower-resolution frames with extra blank pixels to keep the pixel clock rate and total horizontal pixel width constant across modes, allowing the switch to happen without any interruption to the output.

AMD received 1 patent that separates the structural description of a bounding volume hierarchy from the geometry it organizes. In dynamic scenes where objects move between frames, the tree structure itself often remains valid even when positions change, so caching that topology independently allows it to be reused rather than rebuilt from scratch each frame. BVH reconstruction is one of the most computationally expensive steps in real-time ray tracing, and reducing how often it must be fully repeated lowers the overall cost of rendering dynamic environments.

NCSoft received 1 patent addressing the visual awkwardness that can occur when a game character mounts a vehicle or animal. The system uses distance-based triggers to initiate the transition, then blends the character's posture toward the mount using 2 anchor points that connect specific body parts of the player character to corresponding points on the vehicle. Posture interpolation is weighted over time and accounts for the current movement state of both characters, producing a fluid transition rather than a sudden snap between animation states.

Roblox received 1 patent for a creator attribution system that applies branding to virtual assets at the fragment shader level during rasterization, rather than embedding it in the asset file or adding a static overlay. The signature data is stored separately from the asset itself and rendered dynamically by the GPU, adapting to environmental factors like lighting and viewing distance. Because the branding is applied at render time rather than authored into the file, it cannot be removed by modifying the asset.