Performance regarding GPU instancing

Ffigmarion1010 · 3日前

Hello!
I'm using Unity URP and runtime is 4.1
and have question about GPU instancing
1.Does GPU instancing improve performance?
2.If it does, is it improving same object and same animation but different animation time?
(Ex. 2 Monster A with A animation, one monster is playing animation in 32 frame key and other monster is playing same animation at 2 frame key)

3.Material I'm using is Universal Render Pipeline/Spine/Sprite
and lighting mode is Lit Universal.
I want to activate GPU instancing but can't find it.
Is there any way to enable GPU instancing in Sprite material?

I could use spine/skeletion lit material.
but I need to apply emmission so only option is sprite lit universal

Spinebot · 3日前

GPU instancing generally does not improve performance for Spine skeletons because the vertices are animated on the CPU, and the transformed vertices are then uploaded to the GPU. This means that enabling GPU instancing will not provide benefits and might even disable batching, potentially leading to worse performance.
Since GPU instancing does not provide benefits for Spine skeletons, it will not improve performance for the same object and animation with different animation times. The performance impact is more related to how the vertices are processed and uploaded rather than the animation timing.
To enable GPU instancing in a material, you typically need to use a shader that supports it. However, for Spine skeletons, using a non-Spine shader with GPU instancing will not provide benefits. If you need specific features like emission, you might need to modify the shader or use a custom shader that meets your requirements. The Spine/Skeleton Lit material might be a better option if it supports the features you need, but it won't benefit from GPU instancing.

Harald · 2日前

figmarion1010 1.Does GPU instancing improve performance?

No. The skeleton meshes are updated each frame on the CPU as Spinebot said. That's due to the feature set to be supported, it would be too limited if we implemented GPU-skinning for just a feature subset, and then it still would not be suitable for the typical higher vertex count.