CentiLeo 0.650 for Cinema 4D and 3ds Max

The new 0.65 release is all about rendering optimizations compared to older CentiLeo 0.64. Here is a scheme of what is improved:

Base Speedup 

At first we have implemented a base rendering computation speedup that generates the same amount of path tracing samples with the same noise level but with less time. However, this base speedup is variable for different GPU generations:

1. For older GPUs (GTX 1070/1080) the speedup is around 1.5-2x
2. For more recent GPUs (RTX 3070/3080) the speedup is usually 3.5x

The scenes with complex geometry and shaders typically get better render time reduction. For example, the scenes with out-of-core geometry may get up to 5x render time reduction on RTX 3080.

Further Speedup: Render Kernels Option

CentiLeo now has two render kernels in render settings:

1. Bulldozer, intelligent render iteration (the only option of the previous 0.64 version)
2. Motobike, lightweight faster render iteration (the current default option)

The new added option Motobike is a lightweight kernel which in some scenes may accelerate the rendering iteration by up to 1.5x and sometimes with a similar noise level compared to Bulldozer if the lighting setup of the scene is rather simple. However, in different cases (mostly with complex lighting) the Bulldozer kernel may generate significantly lower noise level within the same render iteration and it can be valuable even at higher render iteration time.

Considering our base 3.5x rendering speedup using a new Motobike kernel for some cases we may generate the images with the same number of render iterations and similar noise level up to 5x faster.

Scene Building 

Scene building (compiling ray tracer acceleration structures) is an important part of rendering heavy scenes. We have accelerated the scene builder by around 2.5x. On top of that the second GPU in a system can accelerate the builder of a scene with many million scene primitives by another 1.6x factor. This builder is improved for all scene entities including polygons, particles and tessellations.

Example Interior

Here is the interior render: 32M triangles, 15GB textures, the lighting from outside of room, 2 x RTX3080 10GB VRAM.

Rendering Performance: 2560x1440 image, 32 render iterations (2048 samples), the Motobike render kernel: 24M samples/second, 5:20min for the first image and 5:50 for the second, and then it’s denoised to kill the small residual noise. The Bulldozer kernel renders 32 iterations in 8 min, the noise is lower, but under denoiser it’s killed same way.

Example with Massive Polygon Scene (out-of-core)

Here is the render with huge amount of geometry polygons: 245M triangles, the lighting is HDR + directional light, 2 x RTX3080 10GB VRAM, Motobike render kernel. The amount of polygons is large and only half of them fits to GPU VRAM, so the rendering mode for this scene is out-of-core, but it's as fast as we have a GPU with larger VRAM.

Rendering Performance: 1600x900 image, 78M samples/second, 4 seconds to render 4 render iterations (256 path tracing samples per pixel).

Scene Builder Performance: 12.5 seconds to launch the renderer (read geometry meshes from files and build ray tracing acceleration structures) using two GPUs. The single GPU builds this scene in 22 seconds.

Example with Many Objects and Many Textures

Here is the other scene with a lot of (13600) unique meshes and 12GB textures. 6M unique triangles and 30M instanced. It’s interesting to test loading performance for the scene with so many small meshes.

Rendering Performance: 1600x900 image, 61M samples/second, 6 seconds to render 4 render iterations (256 path tracing samples per pixel).

Scene Builder Performance: one or two GPUs spend 1 second to launch the renderer (read geometry meshes from files and build ray tracing acceleration structures). The previous version took almost 10 seconds to load this scene.

Bulldozer vs Motobike kernel Comparison

Here is a comparison of our 2 rendering kernels for the interior scene from the first example. For both kernels the portals are applied for room windows. On the left image, the Bulldozer kernel executes 4 render iterations in 34 seconds while on the right image the Motobike does so in 19 seconds (1.8x faster) but with more noise.

Multi-GPU Rendering Improvement

In the previous CentiLeo 0.64 there was a problem with the main display GPU that was usually 5-10% slower doing rendering compared to other secondary same GPUs. Now this slowdown of display GPU is eliminated and each element of the multi GPU rig perform the same.