2.77a Vs. 2.78 Cycles GPU render performance

After testing the RC’s of Blender 2.78 I think I can confirm it is slower than 2.77a in the final version.

Environment:

  • blender-2.78-windows64.zip and blender-2.77a-windows64.zip downloaded and unziped in separate folders.
  • bmw27_gpu.blend as the test scene.
  • Just changed system preferences to activate CUDA in both versions.
  • EVGA nVidia GTX 770.

Render time for 2.78 is 8% slower than 2.77a



If you think I did the test the wrong way just let me know.

Can you provide your results to comapre?

Thanks in advance.

Hi. You try the build shared by LazyDodo and do the test again:

Edit:
Buildbot Windows builds seems are back again. You can try those as well.

Tried both options (Latest Builbot & LazyDodo) and results are that they are 5%-8% slower than 2.77a. Bad news I think.

yes… its bad

Just in case to test these Blender 2.78 builds, you try installing the latest nvidia drivers.

Edit:
In my case last builds from buildbot are about 15 seconds faster in BMW27.blend scene compared to 2.77a. But I use Linux, so do not make much sense to compare it with your case.

and CUDA 8 also

I did not install any CUDA for 2.77a. And if I search for CUDA 8 for Windows I only find CUDA Toolkit 8.0 that is supposed for developers.

I have not deep knowledge about this but is it mandatory to install CUDA? I never installed before.

Thanks in advance.

… but is it mandatory to install CUDA?

No, only if you like to compile Blender on your system.
I have tested this too but have same or slightly better performance with latest build.
Specs in my signature.

Cheers, mib

Well then, let’s just install the CUDA 8 toolkit and see what happens.
Without installing CUDA I can’t use CUDA at all, the CUDA version can affect the performance.

@dogdayfear. Have you done the test with the latest nvidia driver?

@cpurender. You do not need Toolkit if you use official Blender including precompiled CUDA kernels. When you install the nvidia driver, it installs the necessary CUDA runtime libraries (libcuda1 package in Ubuntu).
That is, in Windows you need the toolkit only if you’ve compiled Blender for yourself. In Linux you need the Toolkit if you have compiled Blender for yourself, or you are using Blender from Ubuntu repositories (they do not include precompiled CUDA kernels). Install CUDA Toolkit it is not required if you use official Blender, like those from Buildbot.

Installed latest nVidia drivers and results are even worse. (Just 1 second more than the worst result I had).

So I confirm 2.78 is round 8% slower. :frowning: With a few minutes scene render should not seem a problem, but in renders that take hours 8% is too much performance fall.

Try a clean uninstall of the Nvidia driver first.

I don’t have a GTX 770 to try out.
If you can afford a GTX 1070, I would recommend it, it’s much faster than a GTX 770 and consumes less power.

Using the latest builds, I get this on Windows 7:

GPU: inno3D GTX 1070
Tile: 480x540
Hair BVH: off
Render time: 00.52.15

On Ubuntu it’s 3 seconds faster.
https://blenderartists.org/forum/showthread.php?239480-2-7x-Cycles-benchmark-(Updated-BMW)&p=3105909&viewfull=1#post3105909

Hi, make some tests:

[TABLE=“class: grid, width: 500, align: left”]

Master
2.77b

BMW
03:23.02
03:12.85

Fishy Cat
03:25.02
03:30.71

Koro
09:17.27
10:24.30

[/TABLE]

Don´t give to much on BMW, Koro for example is a real production file.
Check Production Benchmark: https://www.blender.org/download/demo-files/
I changed samples/dimension to keep render times low.

Cheers, mib

How many samples/dimensions?

Hi, BMW is default.
Fishy cat is 500 samples
Koro is 200 samples but 100% dimension.
Doesn´t really matter when you compare two Blender versions on one system.

Cheers, mib

Moved from “Latest News” to “Blender and CG Discussions”

I believe there were some excellent tradeoffs between memory usage/features and render time for this release. I know render times are the sexy feature for people playing with Blender in their basement, but memory efficiency (especially on the GPU) and a small tradeoff for faster hair rendering are much more important for studio work so I’d understand why they’re prioritized.

though it would be a whole different story if GSOC denoising gets finished, resulting in less required samples for scenes.

I agree partially if that is the purpose. But consistent improvements involve not sacrifying other features.