hi there and thanks for this writeup! pretty much covers what i've tried/observed so far.
besides setting coolbits to
28 (which i don't think is necessary as a value of
16 should suffice unless one wants to fiddle with GPU-voltages) i've also come across the xorg.conf setting of
Option "RegistryDwords" "PowerMizerEnable=0x1; PerfLevelSrc=0x2222; PowerMizerDefault=0x1; PowerMizerDefaultAC=0x1which ought to help set a performance-level where
GPUMemoryTransferRateOffset can bet set which is in my case only the cards highest performance-level (can be checked with
nvidia-settings -q GPUPerfModes -t)!
however, after booting the card seems to be in the highest perf-level but as soon as i start my miner the perf-leves goes back to the second highest level where the GPUMemoryTransferRateOffset cannot be set.
any idea what i might do wrong on this?

setting fanspeed and GPUGraphicsClockOffset seems to work fine as both are settable not only in the highest perf-level...
driver used on my debian-system is 375.20.
cheers!