Godot performance comparison (master vs 3.x)
The goal of this repository is to compare Godot performance between the
3.x branch and master (upcoming 4.0). This way, we can track significant
regressions and try to address them.
See also godot-benchmarks.
This repository is an independent project because it also needs to run on the
3.x branch, whereas godot-benchmarks makes use of 4.0-only functionality to
provide more independent CPU/GPU timings.
Structure
3.xcontains a Godot project for Godot 3.x (3.3.2 as of writing).mastercontains a Godot project for Godotmasterbranch (unstable).
Both projects should be developed with the same features so that they can be compared with each other.
Benchmark results
- OS: Fedora 34
- GPU: NVIDIA GeForce GTX 1080
- Screen resolution: 2560×1440 (borderless maximized window)
The window manager (KWin) has compositing disabled during the test.
Batching is enabled in both GLES2 and GLES3. Godot's Vulkan renderer does not use batching and makes use of instancing instead.
Note: A debug build is used for both 3.x and master to ensure fairness.
With a fully optimized release build, it's possible to get double the framerate
in 3.x at least (not tested in master since things are currently broken with
optimizations).
Note: The master branch is still in flux, which means the performance
figures here may evolve in the future.
The Vulkan renderer should generally be compared against GLES3 rather than GLES2, since it's much more full-featured than GLES2 (and is also significantly more full-featured than GLES3).
All measurements are in milliseconds per frame rendered (16.66 ms ~= 60 FPS). Lower is better.
2D
| Benchmark | 3.x (GLES2) |
3.x (GLES3) |
Vulkan (master) |
|---|---|---|---|
| 10,000 256×256 Sprites | 20.41 mspf | 40.00 mspf | 23.81 mspf |
| 10,000 4-point Polygon2Ds | 11.63 mspf | 11.36 mspf | 7.30 mspf |
| 10,000 8-point rounded Line2Ds | 62.50 mspf | 62.50 mspf | 8.33 mspf |
3D
| Benchmark | 3.x (GLES2) |
3.x (GLES3) |
Vulkan (master) |
|---|---|---|---|
| 8,000 cubes (1 unshaded material) | 10.20 mspf | 15.87 mspf | 6.94 mspf¹ |
| 8,000 cubes (8,000 unshaded materials) | 20.00 mspf | 20.83 mspf | 6.94 mspf¹ |
| 200 spheres + 200 shadow-less OmniLights | 10.75 mspf | 3.04 mspf | 8.85 mspf |
| 2,000 spheres + 1 shadowed DirectionalLight | 7.35 mspf | 7.94 mspf | 8.62 mspf² |
- ¹: Limited by V-Sync on a 144 Hz monitor. The effective frame time is likely lower than the one reported here.
- ²: Shadow rendering quality is significantly better; it doesn't suffer from acne or peter-panning.
License
Copyright © 2021 Hugo Locurcio and contributors
Unless otherwise specified, files in this repository are licensed under the MIT license. See LICENSE.md for more information.