Call for better official Cycles benchmark

General Forums Latest News
1

It appears that latest hardware from AMD in combination with latest beta
driver gives really good performance on OpenCL rendering, even when using
so called mega kernel. In fact, single Fiji card outperforms double D700
cards from Mac Pro by a factor of 3 here in the studio.

Please follow this link to help benchmark cycles.
Thanks.

lists.blender.org/pipermail/bf-cycles/2015-October/002451.html

2

Did you publish any results already?

Also, I would recommend including a scene of smaller size (in MBytes, but not necessarily less rendering complexity), so the influence of data volume can be taken into account.

3

My friends resuls on mac were super messed up he always needed more than 2x samples to get the same result as cuda or cpu… idk what was going on… however results are identical on windows

4

If AMD cards are finally competitive with CUDA in Blender I would be very interested in good benchmarks to show this. I’m still over here on my old GTX 480, which is great for what it is… but what it is, is about 5 years old. Looking to upgrade early 2016 so this is relevant to my interests.

5

I just hope that AMD performs as expected in compute tasks and improve greatly their Linux driver for games. Right now it is the only thing holding me from buying a R9 290/390/X 8GB.

Sick price/performance wasted in Linux by terrible gaming drivers.

6

so the influence of data volume can be taken into account.so the influence of data volume can be taken into account.
น้ำเต้าปูปลา

7

Edit: 780 Ti results now included.

I do like a good benchmark. Plus I’m about to upgrade (side-grade?) my GTX 590 for a 780 Ti, so am interested to see how the new card performs. I think it should be about as fast, but without the need to disable SLI when using Blender and with a bit more memory.

I had to figure out how to import Sergey’s benchmark files, so this may help others who aren’t sure …

  1. Install TortoiseSVN (I got it here: http://sourceforge.net/projects/tortoisesvn/?source=typ_redirect)
  2. Right click in the folder where you want to put the files and select SVN Checkout
  3. Paste the URL into the top box: https://svn.blender.org/svnroot/bf-blender/trunk/lib/benchmarks/cycles/

The files should now download. So fire up Blender (latest buildbot version) and get rendering!

My results (all CUDA, 256x256 tiles, Win 10/64, 2500k @ 4.9GHz):

GTX 590 GPUx2
classroom: 10m10
fishy cat: no result (out of memory)
koro: no result (out of memory)
pavillon: 12m03
sponza: 7m05

780Ti OC @ 1150MHz
classroom: 8,41
fishy cat: 5m55
koro: 26m53
pavillon: 11m24
sponza: 5m32

Looks like the 780 is a bit faster on these benchmarks, though it is pretty much exactly the same as the 590 on Mike Pan’s BMW benchmark scene. More importantly for me, the extra memory means I can actually render the cat and koro files.

I tried rendering pavillon on OpenCL split kernel just for fun. But it took ages, so I killed it.

Thanks for the benchmark files, Sergey and sparkwoodand21 (if you aren’t the same person).