Which is more efficient using OpenGL's transformation stack or applying the transformations by hand.

I've often heard that you should minimize the number of state transitions in your graphics pipeline. Pushing and popping translation matrices seem like a big change. However, I wonder if the graphics card might be able to more than make up for pipeline hiccup by using its parallel execution hardware to bulk multiply the vertices.

My specific case. I have font rendered to a sprite sheet. The coordinates of each character or a string are calculated and added to a vertex buffer. Now I need to move that string. Would it be better to iterate through the vertex buffer and adjust each of the vertices by hand or temporarily push a new translation matrix?