GIProbe drastically drops performance compared to 3.5 with equivalent settings.

[yosoyfreeman]

Godot version

4.1 stable

System information

Fedora linux, Nvidia 1070 TI

Issue description

Gi probe is exponentially less performant that i's Godot 3x equivalent under similar settings.

Steps to reproduce

I created a basic scene consisting in two CSG nodes to make a simple room. I put a light on the center. In godot 4.0 i made the baking options dynamic to match the 3.5 without the use in bake light option applied and set the resolution to 512 subdivisions. Once baked, i get no noticeable frame drop at all, while in 4.1 the fps tanks to 4 -8 FPS.

Limit size is too low, so i can not upload the baked results, so just press bake in both of them and see the results.

Minimal reproduction project

Test.zip

[mieldepoche]

can confirm the editor starts lagging in 4.1 and not in 3.5.2.
Ubuntu, rtx 3070, nvidia driver 515.105.01

[paskausks]

I don't get really noticable frame drops, but 4.1 does seem to run a bit worse after baking, for me, when compared to 3.5. Also, I got a BSOD after saving the scene for the first time in 4.1, but that's probably unrelated.

Windows 10, NVIDIA GeForce RTX 3080, driver ver 531.79, v4.1.1.stable.official [bd6af8e]

[Calinou]

• Duplicate of Vulkan: MeshInstance3D nodes with Dynamic Bake mode are very expensive in VoxelGI, regardless of quality settings #55359.

In godot 4.0 i made the baking options dynamic

Meshes with dynamic mode are not supported in 3.x GIProbe. Only baked (static) meshes are supported there.

Support for dynamic meshes in VoxelGI is a new feature in 4.0, and it's pretty slow as noted in #55359 as 16× oversampling is currently forced. #55419 adds a way to adjust this oversampling level and also disables oversampling by default, which is much faster.

PS: 512 subdivisions is a lot – you probably don't want to use that in production, especially when using dynamic meshes.