[Bf-cycles] Collecting official benchmark results
sergey.vfx at gmail.com
Fri Oct 30 19:28:28 CET 2015
1) Would say it's more interesting to have statistics about newer cards,
but having statistics about older ones wouldn't hurt at all.
2) That's a valid point. For until recently those benchmarks were only used
by developers to benchmark speed improvements and it makes sense to have
svn version where we can easily do tweaks and fixes. But agree we should
pack those files into easier to download zips. Will do it later.
P.S. Some of the files are available on
http://www.blender.org/download/demo-files/ but those files also has
compositor enabled and such, which makes it so they aren't only measuring
Cycles performance but also postpro performance.
On Fri, Oct 30, 2015 at 10:37 PM, <db4tech at yahoo.co.uk> wrote:
> Hi Sergey,
> Couple of questions
> 1) Are you mainly interested in results from newer graphics cards, NVIDIA
> 700 onwards, rather than older Fermi based cards (GTX 580 for example)?
> 2) Is there a reason the Cycles benchmark downloads do not contain all
> required files in one easy compressed archive (separate archive for each
> scene), consisting of blend file in a relevantly named folder with sub
> 'Texture' folder/s (see below example)? Just an idea that may help prevent
> incorrect results due to a missing texture/s.
> *Classroom.zip example:*
> On 30/10/15 16:52, Sergey Sharybin wrote:
> Hey everyone,
> 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.
> We would like to investigate performance on various hardware and systems
> further, including:
> - CPU performance
> - GPU CUDA performance
> - GPU OpenCL performance
> That's exactly where you can help us! :)
> Here are instructions how do a proper benchmark.
> First of all, you need to use officially approved benchmark scenes from
> . You can grab then using regular SVN command:
> svn co
> Tests are to be done with latest blender builds from our buildbot .
> This is crucial because we like to collect performance from both split and
> mega kernels and switch between them is only possible with latest builds.
> In order to make sure split kernel to use do the following console trick:
> set CYCLES_OPENCL_SPLIT_KERNEL_TEST=1
> Linux / OSX:
> CYCLES_OPENCL_SPLIT_KERNEL_TEST=1 path/to/blender
> In order to make sure mega kernel is used use the same console trick but
> use CYCLES_OPENCL_MEGA_KERNEL_TEST instead of
> Please always do final (F12) renders with the same resolution and number
> of samples as stored in the file.
> Sometimes it's needed to tweak tile size tho, for CPU tile size of 32x32
> is a good setting, for GPU it's 256x256.
> For AMD OpenCL it's recommended to use latest 15.10beta driver.
> When reporting results please include:
> - Exact CPU model you used
> - Exact GPU model you used
> - Timing itself of the scenes you rendered
> - Any modifications you did to the file
> If some configuration failed to render properly it's also interesting to
> For now please submit results here in a reply mail or mail directly to me.
> Once we've got enough data we'll do a nice spreadsheet and share all the
> Thanks in advance,
> Happy Cycling!
>  https://builder.blender.org/download/
> With best regards, Sergey Sharybin
> Bf-cycles mailing listBf-cycles at blender.orghttp://lists.blender.org/mailman/listinfo/bf-cycles
> Bf-cycles mailing list
> Bf-cycles at blender.org
With best regards, Sergey Sharybin
-------------- next part --------------
An HTML attachment was scrubbed...
More information about the Bf-cycles