We recently covered the performance of Turing GPUs in commercial CUDA GPU render applications and OpenCL based render engines. Now it's time to focus our attention on Blender Cycles, the render engine of the free and open 3D creation software. Additionally, we check out the performance of E-Cycles as well.
The Cycles render engine in Blender was assessed with help of the Blender Institute-prepared benchmark pack + two great looking splash screen renders and the recent Barbershop Interior scene from Agent 327 animated feature film. Render time is extracted so that it only covers pure path tracing time (pure dGPU performance) - no kernel compilation, scene loading, CPU-side BVH construction, final composition.
Blender is launched headless (no GUI) with a python script responsible to get everything sorted and start the rendering process. Every render job is repeated three times so that in case of anomaly an investigation can be launched into what went wrong.
You can lower the render times with v2.79b with bigger and padded tile sizes that fit the render resolution of the scene - the benchmark scenes come with quite exotic resolutions and one have to go for different tile sizes for each scene to cover the whole render image with even tiles. Developers made optimizations for v2.80 beta regarding tile sizes and smaller tiles works fine there. One thing worth mentioning that the recent "Creator Ready Driver" (CDR) indeed increases performance in Blender - used a driver from that driver branch that already had those tweaks.
E-Cycles, developed by Mathieu Menuet, is a modified version of Cycles that includes optimizations that significantly lower render times. A short archviz animation sequence is render and an average render time is calculated for a single frame.
It's hard to draw a definite conclusion about the performance of the Turing GPUs in Blender. The "koro" and "fishy cat" scenes with their furry animals see huge speed-up on Turings and v2.80 lower the render times by quite a margin. The "bmw27" shows deltas that you would expect from a brand new uarch. In the "classroom" and "pavillon barcelone" scenes the GTX 1080 Ti really flexes its muscles and is on par with the RTX 2080. Another strong scene for the biggest Pascal is the "Blenderman" scene (Blender 2.78 splash).
It's clear from the results that both the stunning "barbershop interior" scene and the "racing car" scene (Blender 2.77 splash) have some sort of issues on Turing - the RTX 2080 and the Ti variant and the GTX 1660Ti + RTX 2060-2070 producing nearly identical render times.
According to a recently published release schedule, a July launch is planned for Blender v2.80. We benchmarked a buildbot build of v2.80 that was made in middle of March and performance updates arrived since then, so more recent builds can produce faster render times.
The GTX 1080 Ti fares well in E-Cycles with the indoor archviz scenes, it is only slightly slower than the RTX 2080. E-Cycles recieves regular updates and tweaks so performance may have increased since then!
PSU: Cooler Master 1000W VANGUARD
MOTHERBOARD: ASRock X470 Taichi
CPU: AMD Ryzen 7 2700X
GPU: MSI GTX 980 GAMING 4G
GPU: MSI GTX 980Ti GAMING 6G
GPU: MSI GTX 1080 GAMING X+ 8G
GPU: MSI GTX 1080Ti GAMING X 11G
GPU: MSI GTX 1660Ti GAMING X 6G
GPU: MSI RTX 2060 GAMING Z 6G
GPU: MSI RTX 2070 GAMING Z 8G
GPU: MSI RTX 2080 GAMING X TRIO
GPU: NVIDIA RTX 2080Ti FE
OS: Microsoft Windows 10 (10.0) Home 64-bit - Version 1809/RS5 (17763.379)
DRIVER: NV 419.35
RAM: G.Skill FlareX 16GB (2X8GB) DDR4 3200MHz
STORAGE: Samsung 960 EVO NVMe M.2 500GB
COOLER: Wraith Prism with RGB LED