Fastest Render Time

Total Benchmarks

8,296

Benchmarks Last Week

49

Article

Future changes

The last addon update had a very good result. The gpu information is much better now and comparisons do make sense now. I made 2 new overviews to seperate the CPU from de GPU benchmarks. You can find these overview under the Benchmarks menu.

In the meantime I am preparing the next addon update. It will send also the device type with the results, so the website automaticly seperates CPU's from GPU's. I am updating the scene to make it less depending on the CPU when rendering with the GPU. The addon is also going to change the tile size according to the device being used. In the next version the addon will set render settings like samples, resolution and aspect ratio as soon as the benchmark starts. This is to prevent changes made by the user to speed up the benchmark and corrupt the results. When the new version of the addon goes live, the old website will go down. Blenchmark.com is fully operational and there is no use in keeping the old website.

I am also exploring the possibility to create a standalone application using blender. This way also non blender users can benchmark their devices.

This is it for now,
Happy benchmarking!
Mark

Comments

benchmark results inprecise (corrupted) due to variations

Hi,

Thank you for your efforts and contributions!
It is great that you create this standardized method of comparing results.

Yet what I notice is that due the lack of insight in data of the benchmark setup regarding CPU, Motherboard and GPU combinations, the results vary and there for give unreliable results.
(I base this on the assumption that the OS and Blender version is already factored in)

This is caused amongst other due to the following:
- the amount of PCI Express lanes available for the card(s);
- the CPU requires to be able to coop with the GPU's;
- the (amount of) RAM requires to be sufficient to coop with the required data bandwidth consumed by CPU and GPU;

The main problem being that some benchmark results are based on a PCI Express 3 x16 slot, some only use PCI Express 3 x4 for a GFX card.
Which of course may affect the multiple GPU (SLI/Crossfire/..) tests the most.

Even on the same motherboard test results will differ:
>> 4Titan X on the same motherboard:
- ASUS X99-E WS (64 lanes), Intel Core i7-5960X (40 lanes)
- ASUS X99-E WS (64 lanes), Intel Core i7 5820K (28 lanes)

Same motherboard, same GPU's, different CPU, will determine the amount of lanes available.
>> 3 Titan X on the same motherboard:
- ASUS X99-A (40 lanes), Intel Core i7-5960X (40 lanes)
- ASUS X99-A (40 lanes), Intel Core i7 5820K (28 lanes)

Then to count in that the total available RAM (amount & speed) should be sufficient to keep up with the GPU's.
And even the harddisk used will factor in under more extreme circumstances, as someone may test with an old small, slow, low cache, 5400rpm disk in a raid configuration, while others may test with a single, high speed, large cache, SSD.

So it will be VERY INTERESTING if at least the following is taken into account:
- Motherboard, CPU combination;
- More specific: the amount of PCI Express lanes available to each GPU;

And if possible, it would also be interesting to list:
- RAM (amount);
- Harddisc/SSD used.
- (and derived (calculated/estimated) power consumption).

As a final note:
I want to emphasize that tests usually are done with various relatively simple render files. While we attempt to gather reliable data here, we have to take into account that the data is only interesting if we can compare it over time and an as broad as possible range of hardware/software combinations.

It would be more interesting to test not just one file (for instance the BMW file) at low render output, but to include a longer stretch, as the more minutes it takes to render a task, the better you can see the differences in results. With other words, why measure in 10 seconds, as this gives a scale of 10 for the fastest result, while we can measure in 1000 seconds (~17 minutes) with much more precision. Especially for render farms or larger GPU systems the test end results are too close.
Also by testing one file without for instance simulations or particles, we overlook everything, but what is tested.
Perhaps make different tests for:
- large / medium and amount of textures used;
- physics simulation;
- particles and effects;
- mesh render (like the BMW scene);

With four tests, we will get much more reliable results, especially if we can take into account the amount of PCI Express lanes in combination with the CPU used.

Thank you for your time and efforts, I presume many with me, appreciate your contributions, as this will help both Blender development and artists alike.

Thank you very much for that!
I hope you have a great day! :)

Thank you for your comment. I

Thank you for your comment. I will adress this matter in a blogpost later.
I got the same question/request from other blender users as well.

Add new comment

Plain text

  • No HTML tags allowed.
  • Web page addresses and e-mail addresses turn into links automatically.
  • Lines and paragraphs break automatically.

On the Subject of Me

On the Subject of Me

I'm Mark aka Dark, MAD for short. Among other things I am a Blender 3D enthusiast. I founded the Dutch Blender Community forums and build this site to work with the Blenchmark add-on I created for Blender 3D.

Get to know me, see my work

About Blender

Blender is a free and open source 3D animation suite. It supports the entirety of the 3D pipeline‚ÄĒmodeling, rigging, animation, simulation, rendering, compositing and motion tracking, even video editing and game creation.