AMD Patents a Data-Parallel Method to Decompress Dense 3D Triangle Meshes on the GPU

By Patentlyze Team · Updated May 22, 2026

Rendering a detailed 3D scene means feeding a GPU millions of triangle coordinates — and that data has to arrive fast. AMD's new patent describes a way to decompress a compact triangle-mesh format directly on the GPU, without the usual sequential bottlenecks.

What AMD's triangle mesh decompression actually does

Imagine you're downloading a high-resolution 3D map of a city. Every building, road, and lamppost is made of thousands of tiny triangles. Storing and sending all those triangle coordinates takes a lot of space, so engineers compress them — but decompressing them quickly on a GPU has always been awkward.

AMD's patent describes a method that cracks open a particularly compact format called dense geometry compression. Instead of reading triangle data one piece at a time in strict order, the approach uses bitwise and arithmetic operations — the kind of math GPUs do extremely fast in parallel — to figure out exactly which vertex indices make up each triangle.

The trick is that instead of reconstructing every detail from scratch, the method identifies just three things per triangle: the highest index, the lowest index, and a running history of previous codes. From those three anchors, it can fill in all the vertex positions. That keeps the work lightweight enough to scale across the GPU's many parallel cores.

How AMD decodes vertex indices from compressed strips

3D scenes are built from triangle meshes — huge lists of triangles where each triangle is defined by three vertex indices (pointers into a table of actual 3D coordinates). Storing every index explicitly wastes space, so a format called dense geometry compression encodes long strips of connected triangles as compact code sequences, relying on the fact that adjacent triangles usually share vertices.

The patent describes a decompression pipeline that works like this:

• Determine the number of new indices introduced by the current triangle (how many vertices haven't been seen before in the strip).
• Find the highest index referenced by the triangle, derived from the compressed code block.
• Find the lowest index referenced by the triangle, also from the code block.
• Reconstruct all three vertex indices using those two bounds plus a code history — a running record of the compression codes seen so far in the strip.

Critically, these steps are designed to be expressed as bitwise, arithmetic, and logical operations, which means they map naturally onto GPU shader cores. The goal is to let many triangles be decompressed simultaneously rather than serially, turning what was a sequential decode problem into a data-parallel one.

What this means for GPU geometry throughput

GPU rendering pipelines are increasingly bottlenecked by geometry throughput — how quickly triangle data can be fed to the rasterizer. If decompression has to happen sequentially on the CPU before data hits the GPU, you lose the bandwidth advantage of compression in the first place. A method that lets the GPU decompress on-the-fly, in parallel, closes that gap.

This is particularly relevant as games and real-time applications push toward scenes with billions of triangles — terrain, foliage, crowd simulation — where mesh shaders and geometry streaming are already under heavy development across AMD, Nvidia, and Microsoft. A faster, more parallelizable decompression step is the kind of unglamorous infrastructure work that quietly enables higher-quality visuals without requiring bigger memory buses.

Editorial commentary on a publicly published patent application. Not legal advice.