Blender Download Benchmark

[Tarja Hempton]

Todaywe present the Blender Benchmark, a new platform to collect and display the results of hardware and software performance tests. With this benchmark we aim at an optimal comparison between system hardware and installations, and to assist developers to track performance during Blender development.

The benchmark consists of two parts: a downloadable package which runs Blender and renders on several production files, and the Open Data portal on blender.org, where the results will be (optionally) uploaded.

In order to provide control over the data that is shared online, the benchmark result is first associated with the Blender ID of the user, and uploaded on mydata.blender.org, where the user will be able to redact and anonymize the parts containing personal information (Blender ID username and hostname). Currently this information is removed by default. No other personal information is collected.

While hosting Blender Benchmark results will be the initial purpose of the Open Data portal, we plan to host other data sets in the future. For example, information about Blender downloads, telemetry information, etc. Each data set published on the platform will adhere to our Open Data principles, and its collection will be clearly communicated.

The project has been developed by the team at Blender: Brecht van Lommel, Dan MacGrath, Francesco Siddi, Markus Ritberger, Pablo Vazquez, Sybren Stvel and Sergey Sharybin.This project was commissioned by Ton Roosendaal, chairman Blender Foundation.

Thank you for this! Do you perchance know where it calls the blender executable for the timing work? I should clarify, I meant the code for the blender-benchmark executable (on windows this is blender-benchmark.exe).

Downloads are insanely slow!

image641503 82.9 KB

I could download these 200MB within a minute, but at the time I took this screenshot it already spend about thirty minutes and still counting. I actually left my computer alone for the day at this point, hoping to enjoy the benchmark results once back home after work. But nope:

Our connection to the rack is only 1 gigabit, so it can be slow at times due to throttling and QoS. Worse, traffic from Amsterdam to certain places, such as LA or Australia, has been known to be a maximum of about 15mbit, which could be from any number of reasons.

In addition to potentially slow intercontinental bandwidth, we also throttle bandwidth to download.blender.org due to the sheer volume of people we have downloading Blender and the open movies, basically 24/7. Mostly likely this is the limit you are hitting if the speed is limited to around 100KB/sec.

Ideally we could redirect people to mirrors, but this would require a script on our end to dynamically send users to nearby mirrors for blender, as this would not be something you would want to hard code into a binary such as the launcher.

Right, so there are technical limitations, I can understand. Though I still can download from download.blender.org under a minute, or two in rush hours. Is there much of a difference between the benchmark and the website download?

Concerning mirrors, is P2P possible? If the BF provides its official homemade torrents, I bet some people will gladly use them, and it only needs a few scattered around the world to give everyone a great download speed.

Anything is possible! I did some tests of the same torrent software that Ubuntu used and tests went somewhat okay, but there were some tricky issues relating to our internal setup that made it less than ideal, and I eventually threw out the test. Perhaps PTP could be resumed again though.

I believe that Blizzard has been doing this as well for quite a while. I think we also do it for some of our videos for something in Cloud.blender.org or the community site or something, I forget which.

Official blender (from the Arch repos) reports missing libraries (the first is libusd_ms.so which I can only find in blender downloads, though it should be installed from the usd, or at least usd-extra package, but no sign of it).

I can fix that by copying the library to /usr/lib. Then it complains of missing python libraries, and if I fix that, it complains of more and more missing libraries. I gave up, and removed all my hack handed fixes.

Oddly, I then went for blender-git (from AUR), which builds and runs. I had to install optix separately and rebuild the git version to get optix (which is out of date in AUR, but trivial to get the latest version, though I'm pretty sure I've never had to install optix discreetly before).

Blender Institute prepared six Blender files for testing Cycles rendering with CPU/GPU, using various settings and design styles but based on actual production setups. On the links below you can inspect the spreadsheet with results, and load the .blend file collection.

2) Should I use a dedicated GPU for rendering, I mean: connect my displays to a second cheaper GPU, so my main GPU is only used by cycles.

Can anyone tell me if the published developer benchmarks use this setup? I can`t find that information.

What I find most interesting is that any time there is fur or hair in the scene, the 980 and 980 Ti pretty much suck, but the 780 Ti is just stellar. I hope that these results help add some clues to the mystery. All tests were run on 2.77a. My machine specs are posted in field that pops up over the machine name.

Is there a Python script available that runs these benchmarks in sequence, storing the resulting timings (and other stats)? That might make it easier to run the whole set on a varity of our hardware, making sure the settings are the same each time.

Only thing that bothers me is that my GTX 960 is rendering at performance level 3, while it has a level 4 that is not used. My older GTX 660 did use it. I would expect that GPU results could be better at a higher performance level, so what can I do?

So you are a cycle fan ? You have an old computer and graphic card, and want to upgrade your hardware to something good enough for running the new cycle rendering engine ? Here is an interesting article for you.

It all started with a cool model and shared .blend file by Mike Pan. People download the file, render it, and add there rendering times in a spreadsheet. It has been posted a while ago and there have quiet some results (around 160 at the time of writing this post). You can view the original blenderartist thread here :

-2.61-Cycles-render-benchmark

ATI or Nvidia ?

It seems at the moment there is only one viable option for cycle : Nvidia CUDA.

OpenCL is being worked on and hopefully will give some good results soon. For now OpenCl results are also about 20% slower.

Which GPU ?

Well results seem to indicate GTX 570-580 seem good cards, but then it all depends on your budget.

One can note also that Quadro cards do not perform that well. But I believe these cards deliver better performances in display.

It would be much slower to render on a virtual Linux machine. You would have to do a real Linux install to get the benefits. If you want to try it, Ubuntu has some nice utilities that let you carve out a partition for Linux, usually without having to reformat your hard drive. Just make sure to back up you stuff before trying to install a new OS ?

I know that the data is insufficient to answer to questions, but I think it would be interesting to find out how much people payed for their GPU/CPU. That way, we could compare cards based on their time/cost ratio which is a great thing to know if your trying to build a new computer. Who knows, I might look into it if I have time this weekend ?

One thing to add to you post, is that GPU rendering such as cycles does not take advantage of a SLI / Crossfire setup, so getting a matching card is only advantageous if you are also gaming. You can mix and match cards without issue, and Cycles can you both. Just keep in mind that you will be limited by the card with the smallest amount of memory when it comes to scene size.

Nice to see these results but I think a table showing the best laptops for performance might also be of use. Shame the original table did not have an option to select if we were using workstation/laptop, would have made it easier to sort the data.

You would have to use the same system with only the GPU configuration changing to get accurate results. The kind of bloatware someone has in the background can significantly change the numbers. The mother board can also change the results. Especially, with dual or triple cards. For instance, how many PCIe slots on their system are true 16x? Video drivers can have a big affect as well. And many other factors.

The RX 590 is just a GF 12nm respin of the RX 580 which is a respin of RX 480, both on 14nm.

So although this is failing on RX590, a GPU which had a short life, it is very likely failing on other AMD highest market share Polaris GPUs.

The RX 590 cards are both Gigabyte RX 590. One is REV 1.0. the other is REV 2.0.

The REV 2.0 GPU is brand new.

The REV 1.0 GPU is

They both perform equally in 3DMark and there are no problems in gaming on various AAA titles at 2K.

Both cards pass the following standard Blender benchmarks:

Your Radeon Pro WX71000 GPU is Polaris based so I think there must be a generic bug with the Polaris GPUs on Blender.

-graphics/radeon-pro-wx-7100

Since you own a Radeon Pro GPU you should get much better support than people running Commercial GPUs.

Could you file a bug with AMD and get support to get it fixed?

I will see what I can do on Blender Bug Report side and pointy you to a Blender Case Number.

RE: No problems with BMW27, Class Room, Barcelona Pavilion etc, it's "Victor" especific bug ... in my case.

Yes same here. I pass all other tests except "Victor" on the RX590.

But I think "Victor" is the most important, largest, and most realitic test case in the Blender Automated Test Suite.

3a8082e126