So I finally got around to some testing, and used three of my own scenes for comparison.
With CPU only rendering, I noticed little difference in time.
GPU on the other hand, does show in some cases some improvement in a same hardware environment (I have dual GPUs, but for comparison on two of the scenes, rendered using classic rendering on both one and two GPUs since cycles-x doesn’t yet make use of multi-GPU rendering.
On one of my simpler scenes, cycles-x outperformed single GPU classic cycles, but fell behind dual GPU. ON a scene of a character headshot, with hair and SSS, cycles-x outperformed the dual GPU classic render by around 1/3 time.
Scene three, however, is a complex indoor scene, with complex materials. It originally had volumetrics, but I dumped those a few weeks ago for adding rays in the compositor. I deliberately chose this as a test because it is a long rendering scene.
Dual GPU classic renders the scene in 01:19:30.87.
Cycles-X renders the scene in 00:26:06.82. (I put part of the improvement down to lack of caustics in Cycles-X).
Nevertheless, when Cycles-X gets multiple device support, it will blow classic out of the water with no way back.
Yes, there are some differences in lighting when doing a direct side by side comparison, but again, this may be down to caustics.
However, as Cycles-X matures, and particularly the missing features in the roadmap, it will be awesome.
At the moment, Optix render time denoising with Cycles-X is less satisfactory than classic, though OIDN (again, render time rather than post) works just fine, but I expect that will improve, too.