Blender 2.79.5 Cycles GPU Render Comparison with different Tiles (4K)
3d/vfx professor writes:
Most Blender artists are using the official release only and they are used to set higher tiles for faster GPU rendering. Usually by using the Auto Tile Size addon. This is still necessary in 2.79b.
But in the 2.79.5 daily builds there are significant improvements to support optimized hybrid rendering with CPU and GPU. The dev team is also working on multiple CPU/GPU support for single tiles with the goal that tiling doesn’t matter anymore. This means you can use 16×16 for any case.
Note: this isn’t implemented in the latest 2.80 Alpha yet.
In my second comparison I created a LEGO scene with 25,999 LEGO parts and 40,825,145 faces using the principled shader with SSS and textures like scratches and fingerprints.
The comparison shows that rendering at 16×16 is the fastest method.
DELL XPS 15 9560 with GTX 1050 8GB
Blender 2.79.5 c960804
Path Tracing Samples: 128 + Denoiser
26m 58s @ 16×16
31m 04s @ 64×64
31m 01s @ 128×128
30m 52s @ 240×216 (Auto Tile Size)The video has been accelerated to 1,500%.
Awesome! I was right now going to render a picture on cycles when I got the notification of your article! I’m right now trying to render in 16×16 to see how that works.
Thank you so much!
How did it go?
Indeed, the latest daily build version is much faster than the official v2.79b.
I tested the rendering speed of Cycles in pure CPU mode:
v2.79b: 62 seconds
v2.79.0-git20180706.52655f51619-x86_64: 51 seconds
My tests.
On blender 2.79b i had 7:18 (GPU only), on blender-2.79-52655f5-win64-vc14 i had 8:01 (GPU only) and 5:06 CPU & GPU
I’m running blneder 2.79b too, with dual GPU (gtx 970 and gtx690), the 690 itself has 2 GPUs so.. as total i’m using 3GPUs. in the current build of blender, the render supports all GPUs+CPU together in rendering. I didn’t see any difference with larger tile size, but smaller are quite quick in many of renders. like my favorite tile size so far on my this computer (xeon-e5-1660 + 24GB+gtx970+gtx690) is 32×32 or 64×64… it makes slightly better if you go with equal division of resolution.. i.e 1920×1080 => 1920/16 and 1080/16 .
I tried it on the development version 2.79.5 (downloaded on 6-Jul-2018) but the results are like old version.
16×16 is slower than 64×64 which is significantly slower than 512×512
I’ve got an old AMD Octacore and a GTX 960. I’ve tested the new GPU/CPU rendering on different blend files, but most of the time there is no significant speed improvement.
The fastest tile setup seems to be 32 x 32 on my system. But even then, a quite simple image takes 51 seconds to render, the official 2.79b rendering on GPU with 256 x 256 tile size only needs 39 seconds. Using the 2.79.5 build and disabling the CPU, the GPU needs 55 seconds (256 x 256). Which means, the new build is always slower, no matter if using the hybrid rendering or not.
But there’s something where 2.79.5 is really considerably faster on my system: rendering hair. A simple scene took 1:18 minutes (2.79b 1:28), my little Santa Claus model renders in 4:06 minutes on 2.79.5, while 2.79b needs 5:23. That really is an improvement.