High performance out-of-core geometry path tracing demo
Apr 1, 2021 19:46
This demo presents out-of-core path tracing rendering of the Boeing CAD scene with a lot of polygons provided by David Kasik from Boeing Corp many years ago. In this demo we select to render a portion of layers cosisting of 250M unique triangles with high geometry complexity that is visible from every viewpoint. In total the geometry data prepared for rendering including ray tracing acceleration structures occupies 16GB of memory space in RAM. It doesn't enter nor 3ds Max nor Cinema 4D app overloading the RAM too much. So we have made a quick standalone app with CentiLeo rendering core.
And how to render this huge model on a 8GB GPU such as Geforce GTX 1070?
Use CentiLeo GPU renderer with automatic out-of-core technology enabled for such cases. Recall that Windows 10 takes 1.5 GB of GPU memory for own needs, there are also other important CentiLeo system needs like Rays or Texture Cache. The way GPU memory is used by CentiLeo is described here. So we have only 4.6GB left for geometry cache in this case. Actually only 29% of the scene data may fit into this cache or in other words the scene is 3.4x larger than VRAM cache.The complexity of Boeing model comes not only from huge polycount but also from high and complex geometry topology with a lot of small details placed together with larger objects, i.e. the elements are highly irregular.
On the single Geforce GTX 1070, on this old but nice GPU in a 2 bounce path tracing with reflection metallic and diffuse shaders and HDR lighting CentiLeo hits around 10 million path samples per second. It covers a 1600x800 image with 64 samples per pixels in just 10 seconds and updates the viewport interactivelly.
The out-of-core exchange of scene data between VRAM cache and external for GPU RAM storage has 10-15% impact on total measured render time for all view examples from the video demostration. That's an insane caching performance result for the data size 3.4x larger than VRAM geometry cache considering all geometry complexity! And keep in mind, further optimizations for this matter are coming.
Back to the News list