Hi All, I've encountered a regression in power consumption on my SNB mobile GT2 using the latest ~danvet/drm-intel-nightly branch. For example during vaapi playback the power drain jumped from approx 11 Watts to 16 Watts. I've tried to reenable power save state for all components using powertop without effect. I also believe I've found the bad commit: 7499671cb211b34322085172c2fbb33ba87825eb drm/i915: tune the RC6 threshold for stability. It seems that it takes longer for the GPU to switch to a lower C state resulting in a higher average power consumption. Could someone take a look at it and confirm please? Regards Nic
Can you please check with e.g. powertop what the rc6 residency values are for your given workload with and without the offending patch? Just to confirm what's going on. Chris this smells like a candidate for your patches ...
So if you apply http://cgit.freedesktop.org/~ickle/linux-2.6/commit/?h=perf (and build with i915.ko builtin), use xf86-video-intel-2.15+ with SNA, and the lastest checkout of intel-gpu-tools, you get not only a new set of *rps* tweaks but also intel-gpu-overlay to show realtime GPU usage and power consumption. I'd be interested in seeing how the rps tweaks interact with your system (for libva, I can only imagine things get worse). [Not yet integrated into intel-gpu-tools is the ability to record a power usage profile for later analysis - it should be possible through perf, but I haven't finished checking to make sure it works.]
Poke for updated power measurements.
Reporter not responding, closing. Please reopen if the problem persists.
Use of freedesktop.org services, including Bugzilla, is subject to our Code of Conduct. How we collect and use information is described in our Privacy Policy.