What I did was using two modes: in development mode, the shaders would be (re)compiled on the fly (causing some latency), and stored in a bank. In release mode,

What I did was using two modes: in development mode, the shaders would be (re)compiled on the fly (causing some latency), and stored in a bank. In release mode, the shader-bank is used, and a shader compilation would be done if a shader wasnt found - supposedly rarely. As far as I can tell it worked pretty well.