Render resolution and speed

[Asticles]

Hi all

I'm currently working on a scene and I'm seeing big differences in samples/s depending on image resolution, making renders for A3 printing slower unless you use tiled path tracing:

Opencl Path with GPU+CPU Sobol after 3 minutes rendering

Code: Select all

480x270   5.8Ms/s
960x540   5.8Ms/s
1440x810  5.5Ms/s
1920x1080 5.4Ms/s
A3 150dpi
2481x1754 5.1Ms/s
A3 300dpi
4961x3508 3.7Ms/s

Opencl Tiled Path with GPU+CPU after 3 minutes rendering

Code: Select all

480x270   3.1Ms/s
960x540   3.3Ms/s
1440x810  3.3Ms/s
1920x1080 3.4Ms/s
A3 150dpi
2481x1754 3.3Ms/s
A3 300dpi
4961x3508 3.2Ms/s

Is this the correct behaviour?

Thanks

[Dade]

I'm currently working on a scene and I'm seeing big differences in samples/s depending on image resolution, making renders for A3 printing slower unless you use tiled path tracing:

Opencl Path with GPU+CPU Sobol after 3 minutes rendering

Code: Select all

480x270   5.8Ms/s
960x540   5.8Ms/s
1440x810  5.5Ms/s
1920x1080 5.4Ms/s
A3 150dpi
2481x1754 5.1Ms/s
A3 300dpi
4961x3508 3.7Ms/s

Opencl Tiled Path with GPU+CPU after 3 minutes rendering

Code: Select all

480x270   3.1Ms/s
960x540   3.3Ms/s
1440x810  3.3Ms/s
1920x1080 3.4Ms/s
A3 150dpi
2481x1754 3.3Ms/s
A3 300dpi
4961x3508 3.2Ms/s

Is this the correct behaviour?

You probably need to run the rendering with luxconsole (i.e. headless) or to increase the screen refresh time in BlendLuxCore to something like once every 30 minutes. The screen refresh task is probably to heavy at that kind of very high resolution and it is eating your rendering time.

What kind of screen refresh time you are currently using ?

[Asticles]

60s
Yes but look at tiled path speeds.

[Dade]

60s

Try to raise to, at least, 600s

Yes but look at tiled path speeds.

What do you mean ? TilePath is not affected because it works only with tiles and not with the entire image

[Asticles]

Precisely, it is a pity that you lose performance just because the image is bigger.

[Dade]

Precisely, it is a pity that you lose performance just because the image is bigger.

It is currently the only advantage of TilePath over normal Path, leave something the poor fellow

But the difference you have measured it should be due only to the "60secs refresh" setting. There shouldn't be any significate difference with a more prudential setting.

I will do some testing at 4961x3508 here, to check how it works.

Note: your CPU is probably also over-loaded with GPU+CPU rendering and screen refresh. You could also try to reduce the number of threads used by the CPU to have a bit of free space: there is the risk of not being able to feed the GPU fast enough if the CPU has too much load.
For instance, try to leave 1-2 cores free (i.e. with 8 cores, try to use 6 threads to have 2 cores free for no rendering related tasks).

[Asticles]

Ok, I'm trying

[Asticles]

Leaving 2 cores free and setting the refresh interval to 600s I'm getting 4.7 Ms/s after 3 minutes rendering.

Regards

[Dade]

Leaving 2 cores free and setting the refresh interval to 600s I'm getting 4.7 Ms/s after 3 minutes rendering.

Quite better but still not the optimum, I will do some test too.

[Dade]

Quite better but still not the optimum, I will do some test too.

I did some test with i7 3930k (6 cores, 12 threads) + AMD R9 290X + NVIDIA GTX980 with LuxMark Hotel scene at 4961x3508. A screen refresh costs about 2 seconds so it impacts performance but not so much if you do a refresh every 600 secs.

I did some test with NVIDIA and AMD alone without hybrid rendering, CPU 100% free.

NVIDIA GTX980 at 1024x576:

Code: Select all

[Elapsed time:  30/0sec][Samples  100/0][Convergence 0.000000%][Avg. samples/sec  1.97M on 4973.4K tris]

NVIDIA GTX980 at 4961x3508:

Code: Select all

[Elapsed time:  31/0sec][Samples    1/0][Convergence 0.000000%][Avg. samples/sec  0.60M on 4973.4K tris]

AMD R9 290X at 1024x576:

Code: Select all

[Elapsed time:  30/0sec][Samples  120/0][Convergence 0.000000%][Avg. samples/sec  2.33M on 4973.4K tris]

AMD R9 290X at 4961x3508:

Code: Select all

[Elapsed time:  31/0sec][Samples    1/0][Convergence 0.000000%][Avg. samples/sec  0.97M on 4973.4K tris]

So there seems to be a performance degradation at very high resolutions with both. It is a quite strange problem, going to further investigate.