GitHub
Cuda-Based Software Rasterization for Billions of Triangles
Markus Schütz2026-06-06
Open original article →Summary
CuRast is a CUDA-based software rasterizer designed for real-time rendering of massive triangle datasets without precomputed LODs or acceleration structures. It uses a 3-stage pipeline optimized for varying triangle sizes to achieve 2-5x speedups over Vulkan on large dense models. Current limitations include lack of transparency support and inefficiency with many low-poly meshes. Future work aims to add LOD support and transparency.
Key Ideas
CUDA-based rasterization for billions of triangles
3-stage pipeline optimizing small to large triangles
2-5x faster than Vulkan on large models
Focus on dense opaque meshes, no transparency yet
Paper: https://arxiv.org/abs/2604.21749
Related Articles
GitHub
Foxotron: General purpose model viewer
gabrielgambetta.com
Computer Graphics from Scratch
alextardif.com
GPU Tracing Participating Media
imadrahmoune.com
Physically based rendering from first principles
pointersgonewild.com
A Recursive Algorithm to Render Signed Distance Fields
forwardscattering.org