As Anandtech points out: the iPhone 5s' PowerVR Series 6 moves to a fully scalar architecture (allowing better shaders and GPGPU performance) but loses a lot of raw triangle performance in the process (30-50% slower than iPhone 5 in some cases). On shader heavy scenes (T-Rex HD) the tradeoff is worth it (iPhone 5s is 30% faster than any other GPU).

Apple (and PowerVR) have made a couple tradeoffs in this generation that sacrifice performance on some tests for better shader capabilities.

Final point: it'll be interesting to see what these numbers look like when the bugs in the different graphics tests are resolved (are they compiled using the armv8 instruction set?)