Intel Celeron N2820: ACPI CPUfreq vs. P-State Scaling

[phoronix]

Phoronix: Intel Celeron N2820: ACPI CPUfreq vs. P-State Scaling

To complement the recent ACPI CPUfreq vs. Intel P-State Scaling With Linux 3.15 testing that was done using an Intel Core i7 Ivy Bridge Extreme Edition system, here's some similar tests done using a low-power Intel Celeron N2820 "Bay Trail" SoC within the Intel NUC...

http://www.phoronix.com/vr.php?view=MTcwODU

[speculatrix]

I am wondering whether the clever thermal management that Intel use coupled with dynamic clock speed changes mean that "ondemand" might actually be a better policy for people whose computers are idling at least 50% of the time so that the CPU can peak faster before thermal throttling sets in, and thus run better than with max continuous performance settings.

[droidhacker]

How are these figures at all meaningful without a power consumption figure? One can always deactivate frequency scaling and lock everything in at full speed to achieve best performance of all....

[mufasa72]

How are these figures at all meaningful without a power consumption figure? One can always deactivate frequency scaling and lock everything in at full speed to achieve best performance of all....

Agree on first statement, disagree on second. There are actually scenarios in which locking everything a full speed will <decrease> performance, because you don't have the thermal head room to reach the turbo clock rates. Michael has shown this behavior, I just haven't gone to look for the benchmarks.

[Michael]

How are these figures at all meaningful without a power consumption figure? One can always deactivate frequency scaling and lock everything in at full speed to achieve best performance of all....

It was my intention to include power figures, but I only have one WattsUp USB power meter, which has been constantly been for the past week (and counting) with the 50~60+ way graphics card comparison on both open and closed drivers. Thus was unable to obtain accurate power readings for the NUC without it, but after my meter is free again, I may try it again... Considering the situation, that's why these results were just posted as a quick, single-page article for reference purposes.

[toyotabedzrock]

It just occurred to me that their is at least 3 different power save situations and they have different needs.

1. On battery, does the specific battery last longer if the work is done as fast as possible and you allow short spikes in battery usage or is a lower but longer drain better? And does this change as the battery hits 20% charge or lower?

2. Overall power usage. Generally allowing peak performance and allowing the system to return to idle is better.

3. Power save for passively cooled or quiet systems.