[Bf-committers] Couple of questions about rendering settings using python scripts and strange GPU performance

[gandalf3]

Regarding CPU vs GPU, tile size plays a big role. As a general rule of 
thumb, CPUs like smaller tile sizes and  GPUs like larger tile sizes.
It varies from GPU to GPU and from scene to scene, but I tend to use a 
tile size of about 128 for my gpu.

On 04/23/15 22:36, Andrey Chursin wrote:
> Hi
>
> I am processing blender file using python script passed in -P cmd argument
>
> I am new to blender so I am not familiar with all bpy
> objects/properties. Can you help me with it?
>
> 1) How can I get current rendering tile and estimated time in cycles
> rendering engine? I know there is render_stats handler, and I even
> found function BlenderSession::get_progress in sources, but I am still
> puzzled how can I get this info from python script.
> 2) How can I get list of cameras available?
> 3) How can I setup active camera for rendering?
> 4) How can I get/set number of samples when rendering?
>
> And some question about GPU performance
>
> I was rendering first frame from classroom from blender demo
> (http://www.blender.org/wp-content/uploads/2015/04/classroom-200x107.png)
>
> I tested it on c4.large AWS virtual machine (2 cores of Intel Xeon
> E5-2666 v3) vs g2.8xlarge GPU vm(4 unspecified nvidia GPUs with 1,536
> CUDA cores each)
>
> What puzzles me is that 2-core CPU vm renders 2x faster then 4-GPUs machine.
>
> On GPU I was using combined cuda device, and all GPUs was actually
> busy(nvidia-smi shows 99% utilization of all 4 GPUs)
>
> I understand that GPU/CPU are not very comparable, but results are
> still strange. Is there any specific conditions when GPU actually
> works faster then CPU?
>
>


-- 
-gandalf3