This uses GL_ARB_draw_instanced (support > 98.8%) to render some models using instancing instead of individual draw calls.
Essentially, this increases the bandwidth slightly when drawing, since only one draw call is sent for many units. This is dependent on the rendering being somewhat GPU-bound, since logically the CPU has to loop faster than the GPU would render the sent commands.
There is also a bandwidth factor in the matrices that are being sent, but overall I think this is _mostly_ a win.
To actually speed things up substantially, we would need to not re-create the buffer every frame.
Profiling shows a 0-20% perf difference, depending on the map.