[Bf-cycles] CUDA performance tests

David Black db4tech at yahoo.co.uk
Sun May 18 22:55:14 CEST 2014


After several hours of exhaustive tests I'm not certain it is the same 
issue, managed to identify the problem, not the memory increase but the 
reason for the slowdown.

*Tests performed:*
1) To test for the issue you mentioned. Ran several versions of Blender 
with GTX 580m set as Blender display device or purely as GPGPU with an 
HD 3000 driving the display. Result: no noticeable speed variance.

2) Created heavy test .blend (link below) with displacement and hair. 
Result: no noticeable speed variance but evidence of increased memory 
usage with all recent versions of Blender.

3) Removed all textures from my original heavy scene. Result: solved big 
render slowdown but still evidence of increased memory usage.

*Blender versions test**ed:*
2.70a Hash f93bc76 (4th April)
2.70.2 Hash 4ed16bc (17th April)
2.70.4 Hash a8eb95c (27th April) - Memory increase appears in and after 
this version, don't have versions between 17th and 27th to test.
2.70.5 Hash 34bc1e5 VC12 (8th May)
2.70.5 Hash 7891899 VC12 (9th May)
2.70.5 Hash 30361a7 VC12 (13th May)
2.70.5 Hash be980b9 VC12 (16th May)
2.70.5 Hash 2e20c16 (16th May)

*Conclusion**with my heavy scene:*
Windows 7 x64, GTX 580m 2GB.
With Blender 2.70a total memory used 1540, since recent Blender versions 
use more memory (1743) for the same scene there is limited space left 
for render kernel. Removing just one of my high resolution textures also 
works, then total memory used is 1688, it would seem the kernel needs at 
least 300MB spare for full performance? Less than that and performance 
is severely affected, two minute render becomes a ten minute render, 
could a 'Low memory - renders will possibly be slower' warning be 
displayed in the render header?

Can't find a reason for memory usage increase however so supplied test 
.blend (below).

Also noticed. Is it by design that F12 render preview now only updates 
Power of 2 -1 for each consecutive update? This is okay for low sample 
amounts but for renders with a few thousand samples as the render 
progresses it appears nothing is happening or Cycles has frozen and may 
prompt a user to cancel the render.

Memory usage test scene .blend link to compare 2.70a and recent versions.

Thank you,



On 18/05/2014 15:04, Brecht Van Lommel wrote:
> Also, could you mention the commit hash numbers of the builds you
> tested. There were a number of relevant changes on the 17th, so it's
> not clear if they are included or not in that range.
> Thanks,
> Brecht.
> On Sun, May 18, 2014 at 3:57 PM, Brecht Van Lommel
> <brechtvanlommel at pandora.be> wrote:
>> Hi David,
>> Thanks for the tests. I think it's probably the same issue as this bug
>> report, so you can put your findings there:
>> https://developer.blender.org/T40027
>> I'm not sure where the increased memory usage would come from, I can't
>> see any commit in that date range that would explain that.
>> Brecht.
>> On Sat, May 17, 2014 at 7:08 PM, David Black <db4tech at yahoo.co.uk> wrote:
>>> Hi Brecht,
>>> Spent the afternoon carrying out performance tests on a heavy scene (lots of
>>> displacement and hair), sorry scene is test work for a company so unable to
>>> share.
>>> GPGPU tests performed with an 2GB GTX 580m
>>> Small border region render of whole scene (F12 final render)
>>> Blender 2.70a
>>> Time:02:13 Mem:1540
>>> Blender 2.70.4 (and .5)
>>> Time:10:53 Mem:1743
>>> In case there are several accumulating factors involved (other than
>>> __launch_bounds__ ), after testing many Blender versions managed to narrow
>>> down dates, big slowdown and increased memory usage was introduced between
>>> 17th and 27th April.
>>> Also, post slowdown, render preview does not begin until the 12th sample,
>>> thought the render and/or Blender had frozen? Synchronizing objects and BVH
>>> build time is also slightly slower.
>>> Would you like me to open a bug report?
>>> Hope above information is in some way helpful.
>>> David
>>> 3d-designs-davidblack.blogspot.com
>>> On 25/04/2014 16:44, Brecht Van Lommel wrote:
>>> Hi all,
>>> We're having some trouble with CUDA performance at the moment, with in
>>> a performance regression on Titan cards compared to 2.69, and unclear
>>> effects from recent changes. Further there's the new CUDA 6.0 toolkit,
>>> which we ideally could move to, as sticking to 5.0 is a problem with
>>> newer compilers.
>>> We could use some help testing all this, especially for Titan cards as
>>> I don't have one of those.
>>> There's a google doc here:
>>> https://docs.google.com/spreadsheets/d/1IKvO6s7h_0oVGbGE3giowx2fBvyXzIfdM3qCSfROJfI
>>> For now I'm assuming someone with a Titan card can make own builds to
>>> test these git revisions. If not we can build and upload binaries to
>>> test.
>>> The test is:
>>> * download: http://www.pasteall.org/blend/28679
>>> * run: ./blender -b modified_bmw.blend -f 1
>>> If you've got a different .blend file that is giving performance
>>> regressions, or you find different behavior when rendering in the UI
>>> rather than command line, then results from that are welcome too.
>>> The information that I'm looking for from this is especially:
>>> * Which revision caused the Titan performance regressions?
>>> * Is CUDA 6.0 performance acceptable to us for the next release?
>>> * Were there significant performance regressions due to recent changes?
>>> * Which effect do the two provided patches have when building with CUDA 6.0?
>>> Thanks,
>>> Brecht.
>>> _______________________________________________
>>> Bf-cycles mailing list
>>> Bf-cycles at blender.org
>>> http://lists.blender.org/mailman/listinfo/bf-cycles
>>> _______________________________________________
>>> Bf-cycles mailing list
>>> Bf-cycles at blender.org
>>> http://lists.blender.org/mailman/listinfo/bf-cycles
> _______________________________________________
> Bf-cycles mailing list
> Bf-cycles at blender.org
> http://lists.blender.org/mailman/listinfo/bf-cycles

-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://lists.blender.org/pipermail/bf-cycles/attachments/20140518/6915527f/attachment.htm 

More information about the Bf-cycles mailing list