Intel Still Working To Tune Linux Power Efficiency For CPU Power-Sharing Scenarios

Going back to earlier this year has been work on an "adaptive" mode for P-State to improve GPU bound efficiency when the CPU is forced to share a power/thermal budget with other components like onboard graphics. That work is still advancing and an update was provided on it last week.

This work has been under investigation and development for a number of months now for cases where ramping up the CPU frequency in an effort to improve performance may not be as beneficial as remaining in a lower power state so the onboard graphics can enjoy more thermal/power capacity. In some cases this tuning can improve performance by around 15% and with very significant performance per Watt advantages.

But beyond improving the GPU bound efficiency, this work has also led Intel into investigating other possible avenues for improvement with P-State like avoid ramping up the CPU frequency if I/O is the bottleneck and increasing the clock frequency will do little to decrease that bottleneck.

For those interested in enhancing CPU energy efficiency, there is also exploratory work around inferring natural latency constraints from kernel hardware drivers such as based on a monitor refresh rate, audio sample rate, network latency, or other factors in becoming more aware and smarter about whether upping the CPU performance state will be worthwhile for the user.

Those interested in the work going on in this area can see the presentation above as well as the slide decks by Intel power management expert Rafael Wysocki on power budget sharing and by Francisco Jerez of the Intel graphics team with improving CPU energy efficiency during I/O bottlenecks.