Compared to Kepler, Maxwell has been optimized in several details to increase power efficiency. Smaller Streaming Multiprocessors (SMM) with only 128 ALUs (Kepler: 192) and an optimized scheduler should lead to better utilization of the shaders. Nvidia promises that a Maxwell SMM with 128 ALUs can offer 90 percent of the performance of a Kepler SMX with 192 ALUs. GM108 features 3 SMMs and thus 384 shader cores, 24 TMUs and 8 ROPs (64-bit interface).
Another optimization is the massively enlarged L2 cache. The larger size can process some of the memory traffic to allow for a relatively narrow memory interface without significantly reducing performance.
Similar to Fermi and Kepler, the GM107/GM108 support DirectX 12 with feature level 11_0 only.