Thomas Kaiser wrote: > > I have no multimeter that would be precise enough so I rely on the thermal > sensors in the 2 SoCs I used to show the difference the clockspeed makes > depending on dvfs settings.
I bought a powermeter in the meantime and get a difference of ~250mW idle consumption on a Banana Pi M3 (A83T) when switching between interactive and performance (between 2.9/3W with interactive and 3.2W with performance) with the following settings: scaling_max_freq: 1800000 (@1080 mV defined in dvfs table) scaling_min_freq: 480000 (@840 mV defined in dvfs table) Then I defined dvfs settings where all cpufreq operating points share the same 1080mV and tested again. I couldn't measure a difference in consumption realiably but the SoC's internal sensor showed at least some differences. On the left Vcore set statically to 1080mV and on the right the very same test with dynamic voltage frequency scaling and voltage being adjusted between 840mV and 1080mV when walking through _scaling_available_frequencies_: http://linux-sunxi.org/File:A83T_Vcore_static_vs_dvfs.png I still believe that my consumption measurements are too unrealiable to draw any conclusions from. But in the meantime I believe there's some evidence that the whole discussion about power savings or waste of energy due to different cpufreq governors is totally useless if the dvfs settings aren't considered as well. Since based on my tests with A20, H3 and A83T the clockspeed alone wasn't that important regarding consumption and internal SoC temperatures. But the voltage was. Therefore I find it a bit hard to draw _any_ conclusions from these two sets of practical tests that are referenced here since the used dvfs settings aren't mentioned (or I missed it): http://linux-sunxi.org/Cpufreq#The_.22performance.22_governor "The practical tests <https://www.mail-archive.com/linux-sunxi@googlegroups.com/msg00347.html> show <https://www.mail-archive.com/linux-sunxi@googlegroups.com/msg00678.html> that there is not much power consumption difference between idling at 60MHz and idling at 1008MHz (a ~1.5x difference on Cortex-A8, almost no difference on Cortex-A7)" In my tests it made a huge or nearly no difference to switch between both governors -- based on the dvfs settings defined: the higher the voltage difference therein the more difference in consumption. While I still understand that a switch to performance seemed like a good idea two years ago when OS images with 60MHz _scaling_min_freq_ and fantasy or ondemand governor were used, I doubt it's the right approach when taking dvfs settings into account. IMO a better approach is to choose interactive and adjust scaling_min_freq to the cpufreq with the lowest voltage defined in the dvfs table (or maybe one above). At least my tests made with the H3 recently showed nearly no difference whether the H3 idled at 240 or 720 MHz (both defined with the lowest voltage in the dvfs table) therefore I chose 720 as _scaling_min_freq_ and the system behaves as snappy as with performance: http://linux-sunxi.org/File:H3_testing_cpufreq_limits.png I can't test that with an A10 board since I don't own one. But might the test results from back then ("a ~1.5x difference on Cortex-A8, almost no difference on Cortex-A7") not be related more to dvfs settings than A8 vs. A7? Thx, Thomas -- You received this message because you are subscribed to the Google Groups "linux-sunxi" group. To unsubscribe from this group and stop receiving emails from it, send an email to linux-sunxi+unsubscr...@googlegroups.com. For more options, visit https://groups.google.com/d/optout.
