RTX 3090 ~3 times slower than 16-core Ryzen 9950x in a simple scene

[Alex_M]

Hi,

I was testing V-Ray GPU in V-Ray 7 and noticed that it was considerably slower than the CPU engine in a simple scene (screenshots below). Almost 3 minutes with my RTX 3090 vs 1 minute with my 16-core Ryzen 9950x (overclocked to 5.25 Ghz). The results were the same with V-Ray 6, just ~5% slower in both cases. I made sure that the amount of noise is visually the same in both cases to make the test conditions fair. Therefore I had to set the noise threshold for the GPU engine to 0.008 and 0.005 for the CPU engine respectively.



This can't be normal, can it? I was under the impression that RTX 3090 was decently faster than most CPUs, including the 32-core Threadripper 3970x, as reported on a few occasions by Muhammed_Hamed , if I remember correctly.

My V-Ray benchmark CUDA score seems to in line with most other people's scores, so my GPU seems to be working fine.

[Muhammed_Hamed]

Hi Alex,

Other than the bed in your frame, it looks like there is a big empty areas in the frame. V-Ray is able to skip those areas, while V-Ray GPU spends time on empty areas(we have this logged in our system)
Is there a chance you could share this scene with me, muhammed.hamed.ext@chaos.com
We will profile the scene to tell you what is going on

Thanks a lot!
Best,
Muhammed

[Alex_M]

Hi, Muhammed. I've just submitted a support ticket (#312647) and added you as CC. Thanks for looking into this!

[Muhammed_Hamed]

Hi, Muhammed. I've just submitted a support ticket (#312647) and added you as CC. Thanks for looking into this!

Thanks a lot
I will take a look now

Best,
Muhammed

[Muhammed_Hamed]

Thank you for the scene Alex,

My first impression is correct, all pixels around the bed are near empty. The standard V-Ray doesn't waste time on these pixels and focuses on the bed. V-Ray GPU spends as much time on the empty frame as the bed itself, it needs to be smarter to ignore empty pixels. We have this on the to-do list, we will see what could be done about it

Best,
Muhammed

[Alex_M]

Thanks! Looks like I need to be mindful about this until an improvement has been implemented.