Last year I played with clustered deferred shading. The implementation run nicely, and was tested with 1024 real-time lights. The cluster assignment is done with compute shaders, the logarithmic split uses only one variable making the configuration easy, and the shading itself is a simple Phong model.

For this post I rendered the classic Sponza scene without textures but with SSAO that gives the feeling of volume. The SSAO uses 16 samples per pixel in a normal oriented hemisphere, a Gaussian blur, and some tuning to clean out noise.

A video to see it in action, it runs well on an old gtx470 and is not optimized 🙂
(If you have any clue how to encode for youtube better…)

And what I posted on twitter long time ago when I wrote the code :