In the development repository I have just swapped out the old-style projections, which were based on geometric overlap at the grid level, to a quadtree projection method that operates at the cell level. This operation should scale better, should provide much faster runtimes, and is able to operate inline with simulation codes (I have tested this.)
All your projections should now be much, much faster -- on the yt-dev list, speedups of 2-10x were reported before this change was made.
If you choose to upgrade and run into any issues, please let me know or file a bug report. "yt bugreport" can help with this, or see http://hg.enzotools.org/yt/issues/new .