I just finished some extensive testing on standalone VR performance and came to a surprising conclusion.
Right now performance is very similar across all Unity versions. The scene contained quite some tris, but a low amount of materials/draw calls (see screenshot below). Most measurements were within the margin of error.
2020 had slightly better GPU utilization with just opaque materials without any lighting, but as soon as lights and shadows were added 2022.3 and 2023.3 seemed to perform better.
OculusXR is faster than OpenXR (probably due to tech like phase sync, symmetric Projection, and buffer discards). OculusXR did report higher CPU levels, but similar usage. Not sure what to conclude there.
OpenGLES3 is similar to Vulkan.
This means that the base rendering is similar performance-wise, so if you encounter a performance degradation you really need to dig into the URP settings and project settings, reimport packages, and if after that performance still is poor I suggest making a bug report for it.
I probably won’t be doing more tests like this unless I walk into a big performance issue.
The GPU and CPU utilisation is dependent on the level it is running at. 100% utilisation at level 1 may only be 20% utilisation on level 5. So unless you set it to the same level in all your tests, you can’t just compare the utilisation.
Based on your numbers, performance in the first two test cases has significantly improved in the newer versions.
It would be interesting to compare those cases with opengles. This also shows that in the last two cases something happened after 2020 that made CPU performance much worse with the OculusXR plugin.
True. I just tested ‘out of the box’ performance without managing power levels.
In the past it was very hard to get anywhere near this performance in 2022/2023, so now it’s a lot better.
The CPU level was a bit of a weird find indeed. I’ll make a bug report about it and see what they say
2023.3.0a3, just the latest version available.
That issue is soft shadow related, which I haven’t really tried (but if you put soft shadows on low that issue is not relevant either way)
Really appreciate you running these performance tests (and all the bug reports you submit and share around, etc). But as Slin said, with variable GPU/CPU level these results aren’t really valid. The clock frequency differs massively with each level, so the utilization values aren’t comparable when using a different level. The results also don’t mention foveation level, so if dynamic foveation is enabled by default the GPU results will be skewed too.
In this profiling guide, Oculus recommends locking the GPU and CPU levels to 4, disabling dynamic foveation and setting foveation level to 0 via the adb when profiling. See the bottom of this doc for the commands: https://developer.oculus.com/documentation/unity/po-per-frame-gpu/
OpenXR doesn’t utilise foveation as far as I know.
I should have locked the cpu and gpu levels indeed, but 2022 and 2023 getting the same performance with lower gpu levels sounds like a win to me