To begin with an easy test that fully maxes out all of the CPU cores during benchmarking, NASA's NAS Parallel Benchmarks was used while running the Embarrassingly Parallel test-case with the B data-set size (EP.B). As you can see, from the Linux 3.4 kernel to Linux 3.7 kernel there is no real change in performance out of this Intel Core i5 system.
While the performance remained unchanged, the line graph above clearly shows both the Linux 3.6 and 3.7 Git kernels burning through significantly more power than the earlier tested kernels. The newer kernels are going through roughly 20% more power than the earlier kernels when at their peak power consumption with all CPU cores being maxed.
With no change in the raw performance, the overall performance-per-Watt of Linux drops dramatically with Linux 3.6+ for this yet to be resolved problem.