GTX 1080 Ti Various Benchmarks (BMW, Classroom, fishy cat)

General Forums Blender and CG Discussions
41

Welcome to the club. You should see BF 1 with this card. I died constantly because I was just standing still staring at the quality of puddles and grass lol. BTW, Your Blender version is a little old there. Also, the benchmark files I used for these tests are in the first post. Cheers

42

Here are some more render times using today’s build on Ubuntu. I’m going to try bliblubli’s build next on windows 10.

Mar 13 04:08:23 2017 Build

RYZEN 1800X with 2x Rx480, UBUNTU

Demo Files all at default settings, unless noted.

CPU - BMW – 4.27.02
GPU - BMW(2 tiles) – 1.58.71

CPU – Classroom – 13.28.65
GPU – Classroom(2 tiles) – 5.49.22

CPU – Pavillon (32x32) – 11.25.95
GPU – Pavillon (2 tiles) – 1.59.16

43

I know I have the Ti but it is great to see AMD cards getting some blender love. The price / performance is amazing and really good for bringing costs down in the future.

I am just rendering out the tests again using the latest build “e8021f5” https://builder.blender.org/download/ which I will post in an hour or so. Getting lots of different results, also larger tile size does not always equal faster renders (classroom for example).

44

Blender build e8021f5 (13th March)

All GPU, GTX 1080 Ti

BMW Scene
F12 : 2 minutes 21 seconds
4 Tiles: 1 minute 58 seconds
2 Tiles: 1 minute 48 seconds ****
1 Tile : 1 minute 55 seconds

Classroom
F12: 5 minutes 11 seconds (40 tiles) ****
135 Tiles: 6 Minutes 37 seconds
4 Tiles: 5 minutes 35 seconds
2 Tiles: 6 minutes 27 seconds
1 Tile : 8 minutes 11 seconds

Fish Cat
F12: 6 Minutes 33 seconds (1 tile)
32 Tiles: 7 Minutes
15 Tiles: 5 Minutes 39 seconds
8 Tiles: 5 Minutes 16 seconds
6 Tiles: 4 Minutes 45 seconds ****
3 Tiles: 5 minutes 41 seconds
2 Tiles: 5 Minutes 40 seconds

45

If you could rerender the best settings of each scene on the command line with “path/to/blender.exe -b /path/to/file.blend -f 1 --debug-cycles > render.log” to get the pure render time (“without synchronization” it is called), it would allow to really compare times with this thread: https://blenderartists.org/forum/showthread.php?216113-Brecht-s-easter-egg-surprise-Modernizing-shading-and-rendering/page818. Matali took your times which include BVH build times, shader compilation, etc. which depends on CPU and has nothing to do with GPU performance. Lucky enough, for Barcelona it’s one second, so it makes no difference, but for the other ones, it does a pretty big difference.

46

The Issue with doing that is that all other render times in that thread include synchronization. Even though I see your point and agree, we would have to run all the rest of the benchmark test using that method to get an unbiased view of whats really going on.

47

Ok guys, I rerendered everything in UI to get times with synchronisation. Note that the test CPU is at 3.4Ghz with 4 cores. The impact of synchronization may be lower on newer CPUs.

upload.png689×762 7.71 KB


@D3Pixel Maybe you have newer times for Koro? I only found the 2.78c times.

48

Blender 2.78.4 (e8021f5)

Koro
1 Tile: 11 minutes 30 seconds
2 Tiles: 10 minutes 10 seconds ****
6 Tiles: 10 minutes 35 seconds
64x64: 29 minutes 45 seconds

Barcelona
1 Tile: 6 Minutes 31 seconds
2 Tiles: 5 minutes 50 seconds ****
4 Tiles: 5 minutes 44 seconds
6 Tiles: 5 minutes 54 seconds

49

We really need just one Cycles Benchmark file that contains a bit of everything that taxes GPU/CPU. A modern take on the Cornell box so to speak.

A scene that incorporates:

  • Hair & Particle distribution
  • Glossy, Blurry, Reflective
  • Low res HDRI Lighting, high res HDRI reflection (within 3GB limit)
  • SSS on a wedge shaped object
  • Transparent, caustic, Volume Scatter
  • Displacement, Textured and Procedural

This could simply be a 6 x 6 grid of Suzanne heads in a simple environment to catch shadows, lighting, and reflection.

The file should GPU render on 3GB cards?
The file should not take longer than 5 minutes on a higher end GPU
The file should have two preset scenes, one for AMD and one for NVidia for the tile setting difference.
The file should be simple, as in 1. Open 2. Select CPU or GPU Scene from drop down, 3. Press F12
The file should not include any comp effects?
The filename should include the version number
The render should include “stamp output” enabled with date, RenderTime, Frame, Scene, Memory and Filename.
The file should be compatible (if possible) in other render engines like Octane, VRay.

This would make benchmarking Blender quicker and allow for better analysis on the image, as its just one image with everything that a GPU can find tough.

50

The goosberry scene accomplish that, isn´t it?

Maybe except the 5 minute limit I think and the compatibility with other engines, but how other engines work with hair for example makes very difficult that part.

But except those two things I think it qualifies, am I wrong?

Cheers

51

I do not think so, that scene is for CPU because it does not fit in vRAM for most cards.

52

@D3Pixel, interesting proposal. Maybe someone could create an scene that meets the requirements for production, but at the same time it be a simple scene.
The benchmark should be simple to run, people should only use official versions of Blender (buildbot is allowed). One file for CPU and other for GPU. Thinking about the default tiles size that should have the scene for GPU, is a complicated part. Also if the user has Auto Tile Size addon enabled, that will overwrite the default scene tile size.
And there should be some way to get reliable results or validating results to avoid user errors or nvidia/amd fanaticism malicious results. But I think this about reliable results or validation is the most difficult part, it should be something similar to Luxmark.

53

What I am finding difficult at the moment is that each of these benchmark files can render faster with different tile settings, it is very scene dependant. Having just one global benchmark file with fixed samples and a forum friendly resolution might remove the need to constantly re-render with different tlle settings to get faster results. At the moment these benchmarks are all over the place because of this. There can be as much as 30% speed difference (Excluding Koro, something poorly optimised there) just by finding an optimum tile setting which makes these tests less scientific.

I would not have two files (one for cpu, one for gpu) I would have one file but with 3 x “Link Objects” scenes in it.

  1. CPU
  2. AMD
  3. NVidia

This way you just select the scene file from the top of the Blender window. This would also make maintaining the bench file much easier as it is just one file, changes to objects in one of the scenes propagates to the other scenes but the render settings are all separate.

upload.png845×516 49.8 KB

Actually might also need a “4. Hybrid” scene for future when we can have mixed cards + cpu’s all working together.

54

I think that many users do not even know that this field exists and that it is possible to save/choose different scenes there. I think one file per each hardware (like official Blender scenes for benchmark) is the most user friendly. But I do not know…

55

Haha really? People don’t use scenes? It was one of the first things I learnt about after mastering the UI. It is a very helpful Blender feature.

Also, having separate files just for the sake of render settings is inefficient. It’s 3 x maintenance. 3 x the download size etc etc. Duplication is never productive and prone to mistakes. Using linked scenes is more efficient for this purpose regardless if users have never clicked on it before :slight_smile: and it would also teach them something new about Blender if they have never used it…IMO

56

On another note:

Why is Koro 50% slower to render using CUDA compared to OpenCL? (based on latest bench results)

Is hair with transparency something that has been recently optimised on OpenCL?

57

Yes https://git.blender.org/gitweb/gitweb.cgi/blender.git/commit/e8b5a5bf5b63ef1c8980f8da95be32cad4d2cf0e but CUDA too https://git.blender.org/gitweb/gitweb.cgi/blender.git/commit/930186d3df9cb55832d625e96a9655491360ba1e. Each hardware/API has it flaws. And Nvidia didn’t pay a full time dev for a year for Blender development.

58

Thanks for the info BliBluBli

59

@D3Pixel. Believe me, I have seen many users who have never used Scenes.
In addition, for example, if you open a new thread for the benchmark and explain the methodology of the benchmark there, many users will not even read the explanations. They will just open the .blend and start rendering. You see BMW benchmark thread, there are still users who use the old scene.
“cpu”, “gpu_amd”, “gpu_nvidia” name in each .blend file is clearer.

60

Updated the graph with new times, now all times are done on the same build, so comparable. I took the best times for the 1080Ti, so with different tile size. The RX480 always has it’s best times with 64x64 tiles, which is great to spare time, not loosing time searching for each scene what setting is the best :slight_smile:
You can see that koro is now faster on CUDA compared to 2.78c, so the optimization of hair worked.

upload.png763×844 10.2 KB