It breaks down when you approach the hardware limits, i.e. for rts games with thousands of units on the screen, particle effects and so on, or if you have compl

It breaks down when you approach the hardware limits, i.e. for rts games with thousands of units on the screen, particle effects and so on, or if you have complex cad models in automotive applications. Essentially you then need to minimize driver overhead and avoid data copying at all costs. In particular you need to do memory mapped IO (using glMapBufferRange) between the GPU and main memory as fast as possible. That is only possible if you do tripple buffering with a ring buffer, at any given time one part of the buffer will be in flight to the GPU, one will be processed by it and the third will be modified by the CPU. You "publish" with synchronization (glFenceSync() and glClientWaitSync()), see for example http://gdcvault.com/play/1020791/ http://gdcvault.com/play/1020791/. In particular you have to organize all your world data in flat mutable arrays because even on the CPU your data has to be presented in a way that avoids cache misses as much as possible, this philosophy goes by "data-oriented design", see for example https://www.youtube.com/watch?v=rX0ItVEVjHc https://www.youtube.com/watch?v=rX0ItVEVjHc.