One thing I noticed with my i7-10850H CPU is, that if there are cores with huge differences in their frequencies (e.g. one at 900 MHz another at 4.9 GHz) some electrical becomes noticable. The noise can be reduced by limiting the available range the CPUs frequency can be scaled to. Sadly the scaling_min_freq property is ignored outright and only matters as long as the CPU stays idle, but as long as one core gets some load, another one drops below scaling_min_freq, so I couldn't test if limiting the lower bound even helps. Putting one core at max load with stress -c 1 effectively eliminates all noise. This issue is even more annoying on my laptop as it generally runs with the fans turned off. intel_pstate is used in active mode, but putting it in passive mode, disabling HWP or even using the ACPI freq scaling didn't really change anything in this regard. The powersave governor and default energy_performance_preference is used. Is there anything else which could be done to mitigate the problem? Just putting a finger on the touchpad makes the noise quite noticeable as cores get scaled to the max. Thanks
One thing I forgot to mention is, when I limit the max clock, and this just works, the noise gets quieter the smaller the difference between min and max is.
(In reply to Karol Herbst from comment #0) > One thing I noticed with my i7-10850H CPU is, that if there are cores with > huge differences in their frequencies (e.g. one at 900 MHz another at 4.9 > GHz) some electrical becomes noticable. > > The noise can be reduced by limiting the available range the CPUs frequency > can be scaled to. Sadly the scaling_min_freq property is ignored outright > and only matters as long as the CPU stays idle, but as long as one core gets > some load, another one drops below scaling_min_freq, so I couldn't test if > limiting the lower bound even helps. Usually, if we didn't take Pe into account, the CPU frequency should not drop below scaling_min_freq. This sounds like a bug to me. please give turbostat output when this happens. > > Putting one core at max load with stress -c 1 effectively eliminates all > noise. please give turbostat output when this happens. > > This issue is even more annoying on my laptop as it generally runs with the > fans turned off. > > intel_pstate is used in active mode, but putting it in passive mode, > disabling HWP or even using the ACPI freq scaling didn't really change > anything in this regard. The powersave governor and default > energy_performance_preference is used. > > Is there anything else which could be done to mitigate the problem? Just > putting a finger on the touchpad makes the noise quite noticeable as cores > get scaled to the max. > I'd prefer this is the fan noise rather than CPU electrical noise.
Created attachment 294493 [details] turbostat output first three data sets are "idle", next two are finger on the touchpad, last ones are "idle" again. min clock set to 2.4GHz
(In reply to Zhang Rui from comment #2) > (In reply to Karol Herbst from comment #0) > > One thing I noticed with my i7-10850H CPU is, that if there are cores with > > huge differences in their frequencies (e.g. one at 900 MHz another at 4.9 > > GHz) some electrical becomes noticable. > > > > The noise can be reduced by limiting the available range the CPUs frequency > > can be scaled to. Sadly the scaling_min_freq property is ignored outright > > and only matters as long as the CPU stays idle, but as long as one core > gets > > some load, another one drops below scaling_min_freq, so I couldn't test if > > limiting the lower bound even helps. > > Usually, if we didn't take Pe into account, the CPU frequency should not > drop below scaling_min_freq. This sounds like a bug to me. > please give turbostat output when this happens. > > > > > Putting one core at max load with stress -c 1 effectively eliminates all > > noise. > > please give turbostat output when this happens. > attached below. > > > > This issue is even more annoying on my laptop as it generally runs with the > > fans turned off. > > > > intel_pstate is used in active mode, but putting it in passive mode, > > disabling HWP or even using the ACPI freq scaling didn't really change > > anything in this regard. The powersave governor and default > > energy_performance_preference is used. > > > > Is there anything else which could be done to mitigate the problem? Just > > putting a finger on the touchpad makes the noise quite noticeable as cores > > get scaled to the max. > > > I'd prefer this is the fan noise rather than CPU electrical noise. Well, the fans are off, so it has to be something else. It might be something related to the CPU, but the noise it noticeable and doesn't sound like a fan at all.
O_o, that's really weird. I guess this could also be some sort of electromagnetic interference (E.g. with your soundcard? Does the problem occur with the loudspeakers turned off?), or an issue with your PSU. The apparent frequency dropping below scaling_min_freq might be an artifact of how the busy frequency of a CPU thread is calculated: It should approximate the ratio of executed clock cycles to the time it spent in C0 state, which can deviate from its actual working frequency if the processor enters and exits some C1+ state repeatedly since the transition may have some significant latency. According to your turbostat log the CPU cores that report a busy frequency below scaling_min_freq seem to be handling quite a few interrupts which is consistent with that explanation. It doesn't necessarily indicate a power management bug. Does disabling some specific idle state (e.g via /sys/devices/system/cpu/cpu*/cpuidle/state*/disable) have any effect on the noise?
(In reply to Francisco Jerez from comment #5) > O_o, that's really weird. I guess this could also be some sort of > electromagnetic interference (E.g. with your soundcard? Does the problem > occur with the loudspeakers turned off?), or an issue with your PSU. > It is very quiet. Definitely you can hear it only with slowly spinning fans and no background noise. Searching through the web it sounds like a common issue and disabling sleep states is supposed to be the proper "fix" for this. But at least muting the speakers don't help. > The apparent frequency dropping below scaling_min_freq might be an artifact > of how the busy frequency of a CPU thread is calculated: It should > approximate the ratio of executed clock cycles to the time it spent in C0 > state, which can deviate from its actual working frequency if the processor > enters and exits some C1+ state repeatedly since the transition may have > some significant latency. According to your turbostat log the CPU cores > that report a busy frequency below scaling_min_freq seem to be handling > quite a few interrupts which is consistent with that explanation. It > doesn't necessarily indicate a power management bug. > > Does disabling some specific idle state (e.g via > /sys/devices/system/cpu/cpu*/cpuidle/state*/disable) have any effect on the > noise? disabling all except state0 stops the noise.
(In reply to Karol Herbst from comment #6) > > > Does disabling some specific idle state (e.g via > > /sys/devices/system/cpu/cpu*/cpuidle/state*/disable) have any effect on the > > noise? > > disabling all except state0 stops the noise. Ah that makes sense, that sounds like the culprit might be a noisy power supply coupled to the oscillation of the processor's current draw caused by repeated C-state transitions. If that's the case you may be able to tweak the latency parameters to make sure the oscillation frequency is outside the audible range, though I'm not sure if it would make sense to upstream such a change since it will likely come at an energy cost (Though I would expect the cost to be minor as long as the processor can still reach the deepest sleep state in the steady state while idle, might be a good trade-off in your specific case if this is particularly annoying to you).
(In reply to Karol Herbst from comment #3) > Created attachment 294493 [details] > turbostat output > > first three data sets are "idle", next two are finger on the touchpad, last > ones are "idle" again. > > min clock set to 2.4GHz "Sadly the scaling_min_freq property is ignored outright and only matters as long as the CPU stays idle, but as long as one core gets some load, another one drops below scaling_min_freq" I didn't observe this from the turbostat output attached, because the system is idle. Can you please attach the turbostat output when the cpu frequency drops below scaling_min_freq?
(In reply to Zhang Rui from comment #8) > (In reply to Karol Herbst from comment #3) > > Created attachment 294493 [details] > > turbostat output > > > > first three data sets are "idle", next two are finger on the touchpad, last > > ones are "idle" again. > > > > min clock set to 2.4GHz > > "Sadly the scaling_min_freq property is ignored outright and only matters as > long as the CPU stays idle, but as long as one core gets some load, another > one drops below scaling_min_freq" > > I didn't observe this from the turbostat output attached, because the system > is idle. > Can you please attach the turbostat output when the cpu frequency drops > below scaling_min_freq? it does in the output, just only one core per entry, and then only in the entries in the middle. e.g. line 122
(In reply to Francisco Jerez from comment #7) > (In reply to Karol Herbst from comment #6) > > > > > Does disabling some specific idle state (e.g via > > > /sys/devices/system/cpu/cpu*/cpuidle/state*/disable) have any effect on > the > > > noise? > > > > disabling all except state0 stops the noise. > > Ah that makes sense, that sounds like the culprit might be a noisy power > supply coupled to the oscillation of the processor's current draw caused by > repeated C-state transitions. If that's the case you may be able to tweak > the latency parameters to make sure the oscillation frequency is outside the > audible range, though I'm not sure if it would make sense to upstream such a > change since it will likely come at an energy cost (Though I would expect > the cost to be minor as long as the processor can still reach the deepest > sleep state in the steady state while idle, might be a good trade-off in > your specific case if this is particularly annoying to you). yeah... might be. Although it felt like this is a more common issue with newer CPUs, especially 10th gen Intel CPUs and was wondering if there is a way to mitigate it in software. Limiting the max clock helped to reduce it, so I was wondering if there could be some nice ideas for people too lazy or not wanting to bother with replacing the laptop or whatever.
We haven't done such things in kernel before. Compared with the noise, the frequency/performance is one of the indicators that many people care about and are always measuring/monitoring. For people who has a strong requirement of limited noise, they can use the tips you mentioned above. As this is not a software bug, I tends to close it. What do you think?
Bug closed. Please feel free to reopen if you still have any questions.