KeyShot Forum

Technical discussions => General discussion => Topic started by: mikejb on November 10, 2019, 11:47:16 AM

Title: GPU vs. CPU rendering (better?!)
Post by: mikejb on November 10, 2019, 11:47:16 AM
More samples is better, right?

I decided to test KS8 vs. KS9 w/GPU rendering. I loaded in a simple scene and rendered it in exactly the same settings in KS9 and KS8 for 5 mins 30 secs. I was very surprised that KS8 managed more samples than KS9 by almost double:

KS8 5 min 30 sec: 692 samples
KS9 5 min 30 sec: 370 samples

KS9 (on my setup at least) seems to render more samples over time when GPU rendering is turned off. Am I missing something obvious?

My PC Specs:

Intel i7 4ghz
32 gb RAM
Nvidia GTX 1080
Title: Re: GPU vs. CPU rendering (better?!)
Post by: Eugen Fetsch on November 11, 2019, 12:39:41 AM
Is the KS9 number of 370 samples on CPU? What scene have you used?
Title: Re: GPU vs. CPU rendering (better?!)
Post by: DriesV on November 11, 2019, 03:55:02 AM
Hi mikejb

How are you keeping track of samples in KeyShot 9? In the bottom right corner of the output window?
We have identified a bug with GPU Mode that the sample counter, when rendering with Maximum Time, reports incorrectly.
The actual number of samples rendered is higher than what is reported.

This will be fixed in the next update.

Dries
Title: Re: GPU vs. CPU rendering (better?!)
Post by: mikejb on November 11, 2019, 04:36:14 AM
Hello Dries.

Yes, I'm taking note of the samples in the bottom right-hand corner of the render output window. So I assume a better way to test it would be to set the sample rate instead and see if the time between KS9 (GPU) and KS8 is different?
Title: Re: GPU vs. CPU rendering (better?!)
Post by: DriesV on November 11, 2019, 05:10:24 AM
Yes, at least, then the reported samples are correct. :)

Dries
Title: Re: GPU vs. CPU rendering (better?!)
Post by: DriesV on November 11, 2019, 06:07:47 AM
Just a general note...

Please also look at the quality of the rendered images when comparing CPU and GPU output.
We try to make the results as equivalent as possible, but CPU and GPU are still using different algorithms. So comparing speed by comparing render times for a given number of samples is not going to be sufficient.

E.g. Specular HDRI reflections (no roughness) is currently something that requires more samples on GPU than on CPU to resolve cleanly.

Dries