воскресенье, 19 октября 2014 г.

ArchoBenchmark - Blender Cycles render time tests



Blender - 3D software
Cycles - unbiased rendering engine for Blender






Mike Pan's benchmark

Andrew Price's article


Intel Core 2 Duo T6400
Intel Core 2 Duo E6550
Intel i5-3317U
Intel i7-2400K
Intel i7-3770
Intel i7-4700HQ
AMT 1090T


NVidia 550Ti
NVidia 680MX
NVidia 850M


Test 1 scene
Floors, columns and roof - diffuse BSDF
Glass - mix of diffuse, glossy and transparent BSDF shaders
Lamps - emission BSDF, meshes 12 polygons each

Numbers on top - samples. 8x8, 16x16, 32x32,64x64, 128x128, 256x256, 512x512, 1920x1080. Times on top of chart's elements mean minutes:seconds and sometimes hours:minutes:seconds.
Smaller sample sizes lire 8, 16 amd 32 are totally unacceptable for GPUs. Each tile will be rendered very fast, but switching to another tile requires some time too, and these results in terrible render times. Note: robust GPUs suffer even more from such thing.
Overall, GPUs are more effective with big 256 and sometimes 512 tile sizes, while CPUs prefer smaller 16 and 32. But some CPUs show better results with 64 and 128.

Test 2 scene
Made by Mike Pan

Results for 1920x1080 are copied from tests without progressive refine, as render will go the same way - each sample will render the entire image. 1920 sample appears to be useless for CPU. It will be unable to use all of its cores as one sample can be rendered by only one core, therefore render time increases significantly.

For now, Linux and Mac OS X have better render times in comparison to Windows 7 or 8, while tested on the same computer, same file, same render settings and same Blender version.

Progressive refine is obviously slower, because processor needs to switch tiles more, but is useful when the user needs to see the whole image or when it can be necessary to stop render at any time and get a complete picture.

While second test scene is more complex than first, render times increase respectfully for each processor, but dynamics shows that GPUs become more powerful on complex scenes.

Colour and height of chart's element depends on time divided into number samples. More samples means more time, but the dynamics is non-linear, and rendering more samples is more effective. Thus, for example, rendering with 10% more samples will increase render time less than 10%, and on low number of samples this difference can be crucial

Make your contribution

Download files from links in the beginning of this article, send me the results to snowfall.irk@gmail.com, and see your data added to this article and to these charts.