410.66 are useless to me for now. I can also confirm that CUDA is not at fault (the problem is there regardless which version of CUDA I'm using, be it 8, 9.1, 9.2 or 10)
Commands ?
Totally standard Linux stuff for headless:
nvidia-xconfig --silent --cool-bits=28 --custom-edid=HDMI-0:/node/1280x720.bin \
--connected-monitor=HDMI-0 --no-connect-to-acpid --enable-all-gpus \
--allow-empty-initial-configuration --output-xconfig="x.conf"
xinit -- Xorg :1 -config "x.conf" -logfile "$base.log" -quiet &
DISPLAY=:1 nvidia-settings -a "[gpu:0]/GPUGraphicsClockOffset[3]=10
Performance is stock until I execute the above. Then it drops to about 10% of stock, regardless of what I overclock or by how much (even 0).
Something's fishy ...