Matt's Homepage

AMD researchers have published a VRAM-saving technique that leverages procedural generation techniques to eliminate the need for sending the GPU 3D geometry altogether. The GPU utilizes work graphs and mesh nodes to produce 3D-rendered trees on the fly at the LOD (Level of Detail) required for the current frame.

AMD researcher's real-time GPU tree generation system uses work graphs (w/ mesh nodes) for procedural tree generation. Without work graphs, the trees in the scene would have required 34.8 GiB of VRAM. With work graphs, only 51 KiBhttps://t.co/2YcWdOj5Lehttps://t.co/aDkZB08tks

— Compusemble (@compusemble) June 23, 2025

Instead of requiring massive amounts of geometry, the only thing transferred is the code needed to generate the trees in the scene – code that is only a few kilobytes instead of megabytes or even gigabytes.

Read the paper here.

Another article here.

NCOT Technology has a number of retro-programming videos. Bonus points for using old school DOS Borland C++ to compile the examples.

This gives you an idea of how things were done, but the reality is that demo scene code was a LOT more complex – full of inline assembly, crazy lookup tables, direct framebuffer manipulation, and every programming and hardware trick known to man. You had to be a wizard of not only coding – but expert at tricking the hardware to do what you wanted too.

Runevision invented a form of dithering called surface-stable fractal dithering.

He released it as open source along with this video showing how it works:

Find the shader source and example project here: https://github.com/runevision/Dither3D