[Bf-cycles] CUDA Memory usage: Call for testing

[Jeffrey]

Fishy Cat is the only one that could render for me; the rest were 
crashing instantly, but that's for another mail.

Blender master: 5371102
Centos 7 x64
560 Ti 2GB - secondary non-display
970 4GB - display
Nvidia official driver binary 358.16 stable
CUDA TK 6.5
AMD FX-6350 OC to 4.1 GHz

560 Supported: 21:32.37, max GPU memory: 900 MB, reported by blender: 
470.23M
560 Experimental: 22:30.46, max GPU memory: 884 MB, reported by blender: 
470.23M

970 Supported: 12:12.14, max GPU memory: 1548 MB, reported by blender: 
470.23M
970 Experimental: 15:04.27, max GPU memory: 1509 MB, reported by 
blender: 470.23M

The extra time on the 970 experimental kernel is likely because I 
started using my computer a little during the render. I was going out 
when I started the supported kernel render.

On 01/10/2016 11:41 AM, Thomas Dinges wrote:
> Hi everyone,
>
> For Blender 2.72 we added the CUDA Experimental kernel, to have SSS
> working on the GPU.
> The reason we added an extra kernel was mainly an increased memory usage
> during render.
>
> Good news is, this extra memory usage has been fixed by Sergey. I ran
> some tests and rendering several scenes with the Experimental kernel
> results in no or very little extra memory. I only checked GPU memory
> usage (using GPU-Z), not render time.
> http://temp.dingto.org/gpu_experimental_mem.pdf
>
> So please, could people run some tests on their own and report here?
> It's very easy:
> - Download latest build for your OS from the Buildbot
> https://builder.blender.org/download/
> - Render using the GPU, once with the Supported feature set enabled,
> once with the Experimental one.
> - Compare render time and GPU memory usage.
>
> If you compile Blender yourself, make sure you use CUDA Toolkit 6.5 and
> latest master.
>
> You can use scenes from here:
> - Simple scenes:
> https://svn.blender.org/svnroot/bf-blender/trunk/lib/tests/cycles/
> - Larger ones:
> https://svn.blender.org/svnroot/bf-blender/trunk/lib/benchmarks/cycles/
>
> If there is no or little memory / render time increase we can drop the
> extra CUDA kernel, improving Code maintenance and CUDA compilation time.
>
> Thanks,
> Thomas
> _______________________________________________
> Bf-cycles mailing list
> Bf-cycles at blender.org
> http://lists.blender.org/mailman/listinfo/bf-cycles

-- 
Jeffrey "Italic_" Hoover