Ronnie Neeley writes:
I made this video a few weeks ago, but have still yet to see many other people talking about it, so felt it was useful to share again.
GPU + CPU rendering has drastically decreased my render times, and chances are it will do the same for you too!
Viewers were kind enough to share benchmarks. It seems that one needs a mid range to high end CPU to truly benefit from this, but it doesn't hurt to try!
27 Comments
Great news especially if you have an 8-core cpu or two of them. They no longer stand idle.
Very nice. Speed render.
Thanks for this.
So I guess the max memory usable is limited to the GPU's?
I can't really try that feature as I am in production right now and can't change my version of Blender but can you elaborate on the discrepancy in rendering that you mentioned? I sometimes have 2 versions of Blender rendering the same animation, one with CPU one with GPU and I have never seen any difference between the renders. Would be interesting to know if some people have and is it in a specific type of renders (glass, fluids?)
Thanks!
You can test the daily build if you download it as a "portable" version (aka the zip file for windows)(daily builds are shared in portable zip packs anyway) and unpack it aside from your regular blender installation.
Which build is this feature in?
https://builder.blender.org/download/
Windows 64 bit Official blender-2.79-2bf3825-win64.zip
Render hızı PC'ye göre değişiyor. Bende CPU seçiliyken daha hızlı render alıyor. GPU daha yavaş. Benim laptopun ekran kartı eski. CPU 8 çekirdekli daha hızlı.
( İngilizcem çok kötü :) )
Render speed varies by PC. I get faster rendering when CPU is selected. The GPU is slower. My laptop's video card is old. CPU 8-core faster.
(English is very bad :))
3x970 + 1x1060, 3770k, BMW scene:
- GPU, 240x180 - 0:46
- CPU+GPU, 240x180 - 6:36
- CPU+GPU, 64x64 - 1:09
- CPU+GPU, 32x32 - 0:50
- CPU+GPU, 16x16 - 0:42
- CPU+GPU, 12z12 - 0:43
- CPU+GPU, 8x8 - 0:44
Get the Threadripper and be happy.
You have 4 GPU, so your CPU is using only 4 threads for CPU rendering.
In a few tests I've done so far, I've found the new feature to be slightly slower than rendering with my GPU. But what I also found was that render times were drastically cut in the daily build compared to 2.79 stable.
It seems Cycles will hopefully get a lot faster in the future :)
Same here, and no matter what tile size I used the performance is lower,
try 64x64 titles
I am amazed that it has taken this long for someone to get around to coding CPU + GPU rendering. I thought that it was a no-brainer.
I5 6400 @2,7 Ghz + nVidia GeForce GTX 1060 (Windows 10 Pro 64-bit):
GPU 2.78: 01min10.3s
CPU+GPU 2.79 build (32x32): 00min37.38s
Wow! But Blender crashed once.
It's crashing on bigger scenes. Anyway it's promising for the future.
Report sharing a .blend file where the Crash can be reproduced by developers:
https://developer.blender.org/
Under "BF Blender", Report button.
I have a 6 core x2 thread i7 and 2G GTX 960. I'm getting a 30-40% improvement from this feature. Best render times with 64*64 or 128*128 tiles. Great to look forward to, but I'm going to stick to the stable release for now.
i did my compare:
COMPARE 1
MACHINE SETTING:
OSX 10.11.6
GPU - AMD FirePro D700 6144 MB (openCL)
CPU - 2,7 GHz 12-Core Intel Xeon E5
Blender 2.79.1
RENDER SETTING:
Cycles render - 2500AA
CPU - titles 32 - 08:14 min
2xGPU - titles 240/218 - 15:53 min
2xGPU + CPU - titles 240/218 - 07:08 min
2xGPU + CPU - titles 120/108 - 06:22 min
COMPARE 2
MACHINE SETTING:
OSX 10.12 (beta)
GPU - NVIDIA GeForce GTX 780 3072 MB (CUDA)
CPU - 2x 2,66 GHz 6-Core Intel Xeon
Blender 2.79.1
RENDER SETTING:
Cycles render - 2500AA
CPU - titles 32 - 11:18 min
GPU - titles 240/218 - 05:04 min
GPU + CPU - titles 240/218 - 10:05 min
GPU + CPU - titles 120/108 - 05:29 min
A better Cuda card does not have better results. For users with a worse card and a good processor, it makes sense.
my compare here
http://anul147.deviantart.com/art/Compare-CPU-GPU-GPUandCPU-718415896?ga_submit_new=10%3A1512396642
edit:
i did compare GPU+CPU with titles 64x64
COMPARE 1 7:06 min ---- slower
COMPARE 2 4:45 min ---- faster
Result: YES, its faster, than GPU only, BUT don't use Auto title size, find your title size.
I tried to see how rendering with CPU + GPU reacts in a Cuda out of memory scenario.
The answer is, for now, CPU + GPU rendering is still limited by your GPU memory. If ou are out of memory, the render will stop as it already does it with GPU alone.
It woulb be nice if in further version it could automatically switch to CPU render, or even better : keep rendering using RAM and disk memory (like CPU render does) and still using both CPU and GPU render power.
tip: For even faster renders..... Uncheck the Compositor and Sequencer boxes under post-processing, set Antialias 8 samples usually enough, and set type lower than mitchell-Navral(to lowest acceptable quality), under Output set compression to "0" when outputting to .png or .jpeg if file size not an issue... faater render and quality improves as well (you can recompress final video later if needed), (the RGBA setting is only needed for transparent background for compositing, if not needed, use plain RGB to reduce the number of passes per frame) set and watch the render times shrink even more... if you aren't rendering a fleshy or waxy object with subsurface scattering materials, uncheck that setting under Shading settings... After re-importing the still frames as movie clip, Use ffmpeg with mpeg4 with H.264 codec and you'll not only get fast results but still great quality as well.
Hiya!!
Check out Andrew's list:
https://www.youtube.com/watch?v=8gSyEpt4-60
It's from last year, but most of it is still quite relevant...
Cheers,
I'm on LinuxMint 18.3 Cinnamon and that feature works for me only with Blender-Edge 2.79.4, but not with the official 2.79b version. However, it also work with the current 2.8 build.
It only takes one thread of my processor, not all of'em, so the difference is minimal for me.
However also, GPU rendering is significantly faster on Mint than on Windows, which is not the case for CPU, rendering times being virtually the same.
The FX-8350 + RX 480 combo for me takes all the threads of the FX and the shader cores from the RX 480.
Why? The FX-8350 is a fully OpenCL 1.2 compliant CPU.
Funnily, my old i7-965XE is OpenCL compliant, rendering with all threads in CPU only rendering, but taking twice as long to render as via the R9 480X.
I'm currently waiting for the upgrade of Radeon ProRender, promised to be way faster than Nvidia's CUDA.
Has to be tested of course. For what advantage one might ask? Well AMD GPUs are cheaper than Nvidia's and Radeon ProRender way more advanced than CUDA.
OpenCL is the way to go for Cycles like quality, of course...
Eevee has also entered the game, so it really depends on what one is after...
Will rigs with Dual Video cards use both GPUs or just One?