Noticed a similar issue with a E5-1660 on a Sabertooth X79 and Win7.
Currently running 47x100, +112mV turbo voltage offset, default LLC, 0 core voltage offset, 1.2V VCCSA and VTTCPU, DDR3-2133 10-12-12 @1.65V.
Using single threaded IBT as an example here, but any benchmark, application or game that doesn't fully load all cores seems to be affected to some degree.
C-states, C1e, C3, C6, C7 and package C-states enabled, EIST enabled, default "balanced" windows energy settings:
30.6-33.1 GFLOPS, often not seeing more than 4.4GHz max freq, 1.2 idle. HWiNFO shows ~8W CPU package power idle@desktop.
Energy settings min cpu frequency set to 100%:
33.7 GFLOPS, solid 4.7 under load, 1.2 idle. ~9W idle.
C-states disabled, EIST enabled, default energy settings:
23.5 GFLOPS. Runs at 3.3 with single threaded loads (?!?), 1.2 idle. ~24W idle.
C-states disabled, EIST enabled, min cpu frequency 100%:
33.1 GFLOPS. Solid 4.7. ~95W idle.
C-states and EIST disabled, default energy settings:
33.1 GFLOPS. Solid 4.7. ~95W idle.
C-states and EIST disabled, min frequency 100%:
33.1 GFLOPS. Solid 4.7. ~95W idle.
C-states enabled, EIST disabled, default energy settings:
33.7 GFLOPS, 4.7 under load, 1.2 idle. ~9W idle.
C-states enabled, EIST disabled, min frequency 100%:
33.7 GFLOPS, 4.7 under load, 1.2 idle. ~9W idle.
My conclusions:
There appears to be some odd interaction between task migration across cores and Windows CPU frequency scaling.
Disabling C-states massively increases CPU idle power usage and reduces single threaded performance (huh?), so I'd prefer to keep them enabled.
EIST is what enables OS controlled dynamic freq scaling.
So I've settled on running C-states enabled, EIST disabled, default "balanced" energy plan.
Alternatively, C-states and EIST enabled, energy plan min CPU frequency set to 100% would achieve the same result - OS-controlled dynamic frequency scaling neutered while still allowing idle cores to power down.