Arch Linux

tethys

Member

Registered: 2019-08-13

Posts: 170

How is CPU% actually measured?

Normally, when the CPU works at 50% the temperature reaches around 60C on this computer. However, I noticed that when I play some videos, "top" shows CPU% of about 50% but the temperature only rises by about 4C, from 30C to 34C. Certain videos play with low CPU%; the difference seems to be whether the mp4's are "yuv420p(progressive)" in the output of ffmpeg, which play with high CPU%, or "yuv420p(tv, bt709, progressive)", which play with low CPU%. I only noticed this high CPU% / low CPU-temperature in this case. I must mention that I use an nvidia card that, when seen in "nvtop", shares the burden of decoding. But my question here relates to the relation CPU% / CPU-temperature.

Has anyone seen this? Any explanations?

Last edited by tethys (2025-01-27 07:27:20)

Awebb

Member

Registered: 2010-05-06

Posts: 6,688

Re: How is CPU% actually measured?

That percentage is 100% - idle time in percent. Look up CPU governors to understand when and why a CPU might have different clock speeds, but in a  nutshell, 20% at 100 MHz creates less heat than 10% at 3.6 GHz.

tethys

Member

Registered: 2019-08-13

Posts: 170

Re: How is CPU% actually measured?

Thank you for your reply Awebb. Although I could not believe it at first, the 50% CPU at low temperature indeed happen at low CPU frequency, which explains the low temperature of the CPU. I would have expected the CPU% would include the CPU frequency (something like number of instructions per second), but it seems I was wrong.

Awebb

Member

Registered: 2010-05-06

Posts: 6,688

Re: How is CPU% actually measured?

The moment you factor in the frequency, the number gets even harder to interpret. What is the maximum state of the CPU? Is it the multicore max clock or the single core max clock? Is the "computing potential" at the 90% clock/idle coefficient really a linear 90% of whatever a 100% would be? Can you guarantee, that 80% clock/idle are the same at 10° and 40°? Is this even true between to models of the same product series? Between P and E cores? Is the extra overhead of calculating that in real time justified? After all all you do with it, is estimate how busy your CPU is.

I'm not against the concept, I just think that's why it hasn't been done yet.