Simple hybrid mode speed test

[Micha]

Hi,

in the last time I ask me is the hybrid mode worth to think about it. Always if I added my 2950x to the GPU mode I don't got a speed increase.

Today I got the idea for a simple test - I took a scene and rendered it per single 2080ti and than per only 2950x in GPU mode. The result is
2080ti ... 57s
2950x ... 4m53s
So, the 2950x in cuda mode is approx. 5 times slower than a single 2080ti.

Last I have seen the hybrid mode advertisement video and the test hardware was a mobile GPU and a CPU. My test here confirmed my speculation - the hybrid mode is good for users with a fast CPU and a slow GPU only. For users like I who use for example 2x2080ti + 1x1080ti the hybrid mode isn't useful. So, test scene render time per all GPU is 31s. This is ~10 times faster than the 2950x. So, the hybrid mode could allow me to save ~10% of my render time. But in reality the CPU usage needs power which is missed for feeding the GPU and so in hybrid mode my test scene needs 38s. My conclusion for me and my hardware - forget the hybrid mode.

I'm curious - maybe I have overseen something and the hybrid mode is more useful than I think.

-Micha

[Micha]

Same test per V-Ray Benchmark
2950x ... 121
2080ti ... 344

Here the 2950x is only 3 times slower than a 2080ti.

Sidenote: the Benchmark is using nearly 100% of the GPU, but the GPU temperature stay quite low and the reported power consumption is also lower (~190W) than during running my project scene test (~220W). It looks like the real project is harder using the GPU.

[Micha]

One thought more. The 2950x is nearly 5 times slower than a 2080ti in cuda mode. But if I compare the 2950x in CPU mode vs. 2080ti in GPU mode, than I get 2min vs. 1min for my test scene. So, I expect a highest end threadripper like the 2990wx is nearly so fast like a 2080ti. But in hybrid mode the CPU will be slower.

[matanov]

Your conclusions sounds mostly right, except the comparison from the last post - comparing V-Ray and V-Ray GPU will never be accurate because of the tech behind those (hardware and render engine).

[MANUEL_MOUSIOL]

I have the exact same setup and never indulged in any testing.
So this means that all the time when I used hybrid mode, my renderings were actually slower than they could have been?
Is there no way to improve this?

[Micha]

I suppose so simulating CUDA cost a lot of the power of the CPU. In CPU mode the CPU rendering is ok too. I suppose so two threadripper 2990x are quite fast too. But a high end CPU for GPU rendering looks like wasted money if a few GPU are used. So, the question could be, can the CUDA simulation more improved for the CPU? Also the CPU is needed to feed the GPU. Can less CPU power be used here?

[Micha]

So this means that all the time when I used hybrid mode, my renderings were actually slower than they could have been?

Don't trust anything. For example I disabled "adaptive lights" and my render times goes down from 5:16 to 4:48 at my current project.
It's scene with a few lights only and I used adaptive lights 4.

[MANUEL_MOUSIOL]

damn, and I thought I wouldnt need to touch anything anymore. lol
good info!

[Micha]

... and I switched from progressive to bucket mode and got 3:48 now.

[MANUEL_MOUSIOL]

you put my world upside down, Micha!

[francomanko]

so whats the conclusion, hybrid isnt faster than just GPU? I just did a quick test on a scene 1m39 with single gpu (1070) and 1m35 with gpu and i7 cpu. This was in bucket mode...that cant be right surely?

[Micha]

Hybrid helps if you own low GPU power only. If you own highend GPU power than forget hybrid.

There was a video about the speed increase by hybrid mode. If you look here
https://youtu.be/rf1MkuflH3M?t=31
you will see an old single mobile GPU. This card show me on a benchmark side 2.400 points. A 1080ti has 14.000, a 2080ti approx. 17.000 ... .