Here is my new results. Using these settings

Using the same 2000x2180 pix image as before, I get the following:

3090 Cuda: 194 secs, branch tracing off

i9 CPU: 631 sec, branch tracing off

i9 CPU: 218 sec, branch tracing on using 4 rays

3090 GPU Optix: 152 sec

As for visual comparison, I don't see any difference between CPU branch tracing on and off.

Conclusion so far: The render times of CPU with branch tracing on is in the same ballpark as GPU Cude and Optix. GPU is not 5 times faster as in Blender 3.21

So the whole reason for this exercise is to cut the render time for a 1 minute 60 fps lipsyn animation for a customer. A total of 3600 frames.

You can see that the 33 sec Firefly CPU render seems very attractive even though I think the ray-trace render looks better.

Any suggestions to managing the GPU render settings?