AMD has developed a way to reduce the load on video memory by 666 times
AMD engineers have introduced a method for generating 3D trees that virtually eliminates the need for geometry, reducing the amount of occupied video memory from 34,8 GB to all 52 KB. The new technology uses work graphs and mesh nodes, allowing the graphics card to build trees in real time at the desired level of detail - without storing polygons in VRAM.
Instead of traditional geometry GPU receives instructions for generating trees, which take up only a few dozen kilobytes. On the fly, the video card creates the required LOD model using procedural generation. This eliminates the need to load heavy geometry and optimizes streaming scene loading.
The new approach is based on a paradigm in which GPU assigns tasks to himself via work graphs. Mesh nodes allow the video card to make draw calls without the involvement of the processor, meaning the entire rendering process can be performed locally on GPUThis architecture reduces the load on CPU and speeds up content generation.
The technology is not only applicable to trees - a similar approach can be used for other objects, and in the future, perhaps for textures as well. For example, NVIDIA is developing the direction of neural texture compression, and AMD offers a universal solution that does not depend on specific architectures.
