Why Arnold CPU render is Faster than GPU ?

[yingzhiworkspace]

I had an I7 14700k, 64 GB DDR4 3200mhz RAM, and a 3060TI 8G setup. While testing out for some project, I compared the render time on CPU and GPU, which was hugely different and produced more noise on GPU.

Scene A Setup:
A Static Camera with Animation ,  with a lot of props , main building and subject scatter, 80% diffuse shader , 5% Tranmission Shader and 15% metallic Shader

Scene Complexity : Mid-high

CPU Render Time Average per Frame :  23s
CPU Sample AA  Setting: 4/3/3/3/0/0 (No Adaptive enable)
CPU Noise Level : Mild (subtle on AO area , rest is fine , no denoiser)

GPU Render Time Average per Frame : 3.3 mins 
GPU Camera AA : 4 , Adaptive Sample : 25 , Thresold : 0.015
GPU Noise Level : High specially reflective and tranmission surface , Sequence blobby on glass shader .

Light Source : only hdri with 2 sample (Same setting on CPU and GPU)

Scene B Setup:
A cloner with custom mesh and animate form into a radial shape, Only Metallic material applied.

Scene Complexity : Easy

CPU Render Time Average per Frame :  Min : 15s / Max : 2.5mins
CPU Sample AA  Setting: 5/3/5/0/0/0 (No Adaptive enable)
CPU Noise Level : Mild (subtle on AO area , rest is fine)

GPU Render Time Average per Frame : Min : 1.5 mins / Max : 4 mins
GPU Camera AA : 4 , Adaptive Sample : 25 , Thresold : 0.015
GPU Noise Level : High specially reflective surface , Sequence blobby when using denoiser.

Light Source : only hdri with 4 sample (Same setting on CPU and GPU)

Maybe I'm wrong with the settings, but I couldn't find much more to improve rendering speed on GPU. Anyone advise or suggest on this?  Thank you

[Stephen.Blair]

Isn't that one of the problem CPUs? That's going to cause problems even with GPU.

Full logs would be great, or even a test scene.

[yingzhiworkspace]

Hi @Stephen.Blair thanks for the reply, here's attach with the test scene, HDRI is not included, you may grab any Studio HDRI from HDRI Haven and give a try. Very appreciate for the help ! The issue right now was GPU produce more noise and render much more slower than CPU, Thank you.

[yingzhiworkspace]

OK after I test it out for few scenario, I do realise threshold for adaptive sampling on GPU is not necessary always to be at 0.015 or lower for high quality output, as sometime even 0.2 will do a great job on cleaing up till acceptable noise level and improving render speed, unless that is real hard to clean than only go for a lower thresold value. Correct me if i am wrong. Thank you