Currently motorcar clients (clients creating 3D windows) pack their depth buffer into their color buffer and then the compositor unpacks this pack into a depth buffer for use during depth compositing (see this slide for details). This adds an extra render pass on the client side and the compositor side, which entails rgb8<->float32 conversion per pixel per client per frame across the entire framebuffer, which is fairly expensive.
This issue tracks investigation of using a separate float32 buffer for the depth buffer (if possible)