https://hgpu.org/?p=3619
Data handling inefficiencies between CUDA, 3D rendering, and system memory