RTX 2070 slower than GTX 1080 TI in barber shop scene?


Does anybody know why demo barber shop scene rendering (https://www.blender.org/download/demo-files/ ) is slower using a RTX 2070 (17m 34s) than a GTX 1080 TI (9min 2s) ?

If I use optix in the RTX it takes even longer. RTX is faster in other scenes as expected, though.

I have a 1080 and a 2070 super. Here are my results with latest alpha:
1080: 12m 51s
2070s: 17m 42s

1080: 38m 38s
2070s: 09m 32s

But results vary heavily between different releases in this scene and optix seems to be sometimes slower then cuda. The good thing about the optix render, it’s almost noise free!

You can also use Optix denoiser in CUDA renders

@Feraud , You can filter results by scene, hardware and device type here to compare times with yours:

I am not talking about the denoiser. The render (without denoising) in optix is almost noise free.
Like Feraud said, the barbershop scene is different in render times compared to other scenes, like the benchmark (Blender - Open Data). Somehow the 2070 is slower or similar to gtx cards and optix is really slow depending on the blender build.

Could you share a scene where that happens? That is, with exactly the same settings in the scene, OptiX and CUDA render results being different in terms of noise.

As I understand it, any device should give the same result.

rendered with 1080 in cuda (12m41s)
rendered with 1080 in optix (38m58s)
No settings changed. Same render results with the 2070 Super (Cuda: 11m51s, Optix 17m14s).
It’s the barpershop scene: https://svn.blender.org/svnroot/bf-blender/trunk/lib/benchmarks/cycles/barbershop_interior/

Just found out that branched path tracing is not supported in optix: https://developer.blender.org/D6202
It’s silently switching to path tracing with different settings. That explains the difference between cuda and optix.

1 Like

That might as well be the answer for you too, @Feraud

Basically, you are comparing two device types with different settings in the scenes. You switch to Path Tracing, you save the scene with a different name and you then do the render time comparisons again.
I mean, when you use OptiX