(Test) Particle Rendering: Point Sprites vs Geometry Instancing based Billboards

In this article, I compared two methods for rendering a lot of particles: 1 – the first method is based on point sprites GPU built-in feature. We send an array of points, and the GPU transform them into billboards (quads facing the camera). 2 – the second method is based on the first: we send an array of points, but we use a geometry shader to create billboards. On AMD Radeon GPUs, geometry shader sprites are 30% slower than point sprites. On NVIDIA GPUs (GTX 780, GTX 680), geometry shader sprites have more or less the same performance than point sprites.

Now let’s see a third way, this time, based on geometry instancing: we have a quad (4 vertices) and we use the GPU built-in geometry instancing feature (GL_ARB_draw_instanced + GL_ARB_instanced_arrays, OpenGL 3.3+) to render a lot of quads, each quad being a particle. The billboarding is done for every instance of the quad in the vertex shader.

A demo, coded for GLSL Hacker, is available in the host_api/Particle_Geometry_Instancing/ folder of the code sample pack. Start you favorite FPS-meter (FRAPS, etc…) and load the particle_gi_1920x1080_fullscreen.xml file in GLSL Hacker.

Testbed:
– CPU: Intel Core i5-4670K @ 3.4GHz
– Mobo: GIGABYTE G1.Sniper M5
– Memory: G-Skill 16GB DDR3 1600MHz
– Windows 8 64-bit
– Catalyst 14.4 (for Radeon cards)
– R344.11 (for GeForce cards)
– GLSL Hacker 0.7.0.4
– FRAPS for displaying the framerate in fullscreen mode.

Settings: particules: 1’000’000, resolution: 1920×1080 fullscreen.

	Point Sprite (PS)	Geometry Shader (GS)	Geometry Instancing (GI)	Difference (GS/PS) | (GI/PS)
Radeon HD 7970	147 FPS	105 FPS	134 FPS	-28% | -8%
Radeon HD 6970	98 FPS	65 FPS	91 FPS	-33% | -7%
Radeon HD 5870	85 FPS	59 FPS	77 FPS	-30% | -9%
GeForce GTX 780	295 FPS	279 FPS	97 FPS	-5% | -67%
GeForce GTX 750	203 FPS	122 FPS	104 FPS	-39% | -48%
GeForce GTX 680	72 FPS	71 FPS	73 FPS	-1% | +1%

On AMD Radeon GPUs, geometry instancing work rather fine for particle rendering. This method is slower than point sprites but the difference is smaller compared to geometry shader method: geometry instancing is 10% slower while geometry shader is 30% slower.

On NVIDIA GPUs, results are a bit difficult to interpret… Clearly, rendering particles with geometry instancing is not the strength of GeForce cards. Almost 70% slower on the GTX 780! The GTX 780 (Kepler) is even slower than the GTX 750 (Maxwell first gen)! Only the GTX 680 remains constant. I think I will do a geometry instancing test to check how Radeon and GeForce behave when the number of triangles per instance vary, just to be sure that these results are coherent.

Conclusion: Point sprites are the fastest way to render a lot of particles. If you need more control over particle rendering, you can use geometry instancing with Radeon GPUs and geometry shaders with GeForce GPUs.