Cpufreq and overclocking (bug)

Reasons · 2008-07-30 21:35:05

While I started a thread on it, (link) I'm thinking this might be a bug. On a processor overclocked, modprobing acpi-cpufreq sets the speed to stock speeds and doesn't scale. If the option for intel speedstepping is disabled in the BIOS then it won't modprobe at all.

Processor: Intel q6600
Moterboard: ASUS PKK-E
Arch: x86_64

jleach · 2008-07-30 22:39:08

Overclocked Q6600 here on Arch x86_64, acpi-cpufreq module seems to be scaling correctly.

I think your issue may be that the default setting for the 2.6.26-ARCH stock kernel is the "performance" governor, which I believe will always run the cores at maximum speed. Try setting it to "ondemand" or "conservative":

for i in `seq 0 3`; do cpufreq-set -c $i -g ondemand; done

Then check it with cpufreq-info, you should see some or all cores will drop down to lower speeds when the computer load is low.

Reasons · 2008-07-30 23:03:42

jleach wrote:

Overclocked Q6600 here on Arch x86_64, acpi-cpufreq module seems to be scaling correctly.
I think your issue may be that the default setting for the 2.6.26-ARCH stock kernel is the "performance" governor, which I believe will always run the cores at maximum speed. Try setting it to "ondemand" or "conservative":
for i in `seq 0 3`; do cpufreq-set -c $i -g ondemand; done
Then check it with cpufreq-info, you should see some or all cores will drop down to lower speeds when the computer load is low.

That's the problem though; cpufreq-info shows the following which limits everything at 2.39 instead of the 3.6 I have it clocked to. It will boot with that speed (shown in conky) until I modprobe it.

cpufrequtils 004: cpufreq-info (C) Dominik Brodowski 2004-2006
Report errors and bugs to cpufreq@lists.linux.org.uk, please.
analyzing CPU 0:
  driver: acpi-cpufreq
  CPUs which need to switch frequency at the same time: 0 1 2 3
  hardware limits: 1.60 GHz - 2.39 GHz
  available frequency steps: 2.39 GHz, 1.60 GHz
  available cpufreq governors: ondemand, performance
  current policy: frequency should be within 1.60 GHz and 2.39 GHz.
                  The governor "performance" may decide which speed to use
                  within this range.
  current CPU frequency is 2.39 GHz.
analyzing CPU 1:
  driver: acpi-cpufreq
  CPUs which need to switch frequency at the same time: 0 1 2 3
  hardware limits: 1.60 GHz - 2.39 GHz
  available frequency steps: 2.39 GHz, 1.60 GHz
  available cpufreq governors: ondemand, performance
  current policy: frequency should be within 1.60 GHz and 2.39 GHz.
                  The governor "performance" may decide which speed to use
                  within this range.
  current CPU frequency is 2.39 GHz.
analyzing CPU 2:
  driver: acpi-cpufreq
  CPUs which need to switch frequency at the same time: 0 1 2 3
  hardware limits: 1.60 GHz - 2.39 GHz
  available frequency steps: 2.39 GHz, 1.60 GHz
  available cpufreq governors: ondemand, performance
  current policy: frequency should be within 1.60 GHz and 2.39 GHz.
                  The governor "performance" may decide which speed to use
                  within this range.
  current CPU frequency is 2.39 GHz.
analyzing CPU 3:
  driver: acpi-cpufreq
  CPUs which need to switch frequency at the same time: 0 1 2 3
  hardware limits: 1.60 GHz - 2.39 GHz
  available frequency steps: 2.39 GHz, 1.60 GHz
  available cpufreq governors: ondemand, performance
  current policy: frequency should be within 1.60 GHz and 2.39 GHz.
                  The governor "performance" may decide which speed to use
                  within this range.
  current CPU frequency is 2.39 GHz.

jleach · 2008-07-31 00:12:06

Pretty strange, but I'd be thinking it's more related to the motherboard BIOS (updated to your latest?) and ACPI support (does 'dmesg | grep ACPI' reveal any warnings or errors?).

See also:
http://bugzilla.kernel.org/show_bug.cgi?id=8950
http://bbs.archlinux.org/viewtopic.php? … 17#p351117.

I'm also running at 3.6GHz, but I'm using an Abit IP35 Pro motherboard:

hardware limits: 2.40 GHz - 3.60 GHz
available frequency steps: 3.60 GHz, 2.40 GHz

I couldn't even find any web info for the Asus PKK-E (actually this thread comes up first on google )

Reasons · 2008-07-31 01:45:24

Should have been P5K-E. No errors in Dmesg. I'm rather fond of the BIOS version I have now but I'll try to find a newer one. I wouldn't think it'd be a big problem as it's an ASUS and they're great with linux.

Reasons · 2008-08-01 23:07:35

The BIOS is the latest revision and the problem still continues.

Reasons · 2008-08-02 20:30:25

I'm going to try compiling kernel 2.6.26 to see if it helps.

litemotiv · 2008-08-03 21:43:56

i opened a bugreported for this on launchpad about a year ago:

https://bugs.launchpad.net/ubuntu/+sour … bug/132403

and e-mailed dominik brodowski (cpufreq), who claimed it's a bios/motherboard bug reporting the wrong acpi-table values to linux. scaling my overclocked c2d worked perfectly fine on windows xp, brodowski however said windows uses workarounds which aren't conform 'standards'.

JGC · 2008-08-03 22:05:54

Doesn't speedstepping just change voltage and multiplier? I haven't seen acpi-cpufreq changing FSB, if that would be possible, a C2D CPU would have much more possible speedstep settings.
I guess your BIOS doesn't give the real speeds for your speedstep settings, but the speeds that would be when you're not running overclocked. Try something like this:
- boot CPU at stock speed
- set cpufreq at lowest setting
- run some benchmark
- reboot, set CPU overclocked
- set cpufreq at lowest setting again
- run same benchmark

If I'm right about this, your CPU will be overclocked even when using speedstepping, while showing the not-overclocked frequencies.

These features differ with each and every mainboard. With my Gigabyte mainboard, as soon as I overclock the FSB with just one MHz, I can't modprobe acpi-cpufreq without error messages anymore.

Reasons · 2008-08-03 22:07:42

litemotiv wrote:

i opened a bugreported for this on launchpad about a year ago:
https://bugs.launchpad.net/ubuntu/+sour … bug/132403
and e-mailed dominik brodowski (cpufreq), who claimed it's a bios/motherboard bug reporting the wrong acpi-table values to linux. scaling my overclocked c2d worked perfectly fine on windows xp, brodowski however said windows uses workarounds which aren't conform 'standards'.

That's nice to know, and slightly annoying at the same time. I already did benchmarking and really it's linux that's not taking the power it can.

.:B:. · 2008-08-03 22:09:40

Why do you want to do power saving on an overclocked CPU? A bit of a contradictio in terminis no?

Reasons · 2008-08-03 22:18:10

B wrote:

Why do you want to do power saving on an overclocked CPU? A bit of a contradictio in terminis no?

I don't hate the world enough to want it to run at 3600MHz all the time, 2400MHz is more than enough to idle at.

litemotiv · 2008-08-03 23:38:23

@JGC: yes, that's what Brodowski suspected too, he didn't want to look into the issue further.

Reasons wrote:

That's nice to know, and slightly annoying at the same time. I already did benchmarking and really it's linux that's not taking the power it can.

yes i benchmarked it too back then, and also concluded it was more than a 'labelling' problem. the simplest test i did was setting the governor to userspace and forcing it to 3ghz (my overclocked speed). lm_sensors gave that back as 2.39ghz (stock). then i rmmodded acpi-cpufreq and lm_sensors printed 3ghz, the temperature went up as well as the fanspeed.

tuxfusion · 2009-10-17 14:07:36

I know it is old, but still unsolved.

I overclock a def. 3G e8400 to 4G

Funnily my systems shows only /goes up to 3,6 with acpi-cpufreq no matter what i try , including
manipulating /sys/devices/system/cpu/cpu0/cpufreq/cpuinfo_max_freq.

Benches are fast, but when i remove acpi-cpufreq , freq jumps back to 4G
and i get far slower benchmarks !

I suspect 3,6 is just a wrong reading ( by bios ) this would explain the faster benches.

I wanted to throw in , that dmidecode reads the correct value FSB + MHZ.
However it does not update when cpufreq catches in ...

graysky · 2009-10-17 14:52:44

I'm using an overclocked X3360 just fine w/ powernowd instead of acpi-cpufreq and it bounces from 2.40 GHz to 3.40 GHz just fine. You can get powernowd from the AUR and get it up and running in no time. It adjusts the speed of your CPU depending on system load. It is more flexible than the ondemand profile in cpufrequtils because you set the low/high thresholds of CPU use to trigger a step from low multipler to high multiplier rather than relying on the 'ondemand' profile.

http://wiki.archlinux.org/index.php/Powernowd

...you know what the ironic part about speedstep is? Stepping from 6x400 to 8.5x400 doesn't do anything for power consumption as measured on my Kill-A-Watt

Last edited by graysky (2009-10-17 14:53:52)

Dehir · 2010-05-09 08:42:39

Thou this is an old thread. But im still having similar problem with my e8400 c2d and asus p5k-pro mb. With automatic and default clockspeed my cpu scales down to 2Ghz but when i set manually multiplier and overclock my cpu and memory to 3.6ghz it wont load acpi-cpufreq module. Currently running kernel 2.6.33.

But recent research i found that when i set my mb to overclock profile it drops speedstep option away from cpu settings at bios. So there are no actually even possibility to have speedstepping while overclocked by this mb/bios. Which really sucks. As it appears to me correct me if i am wrong.

Last edited by Dehir (2010-05-09 08:53:54)

jarda-wien · 2010-09-27 09:51:44

I have an ASUS P5K64 WS with BIOS 0801.
First of all, there is no speedstep option in the bios, but I have figured out, that some options enable/disable it (it is quite logical btw as you will see...).
What cpufreq does is that it lowers the CPU multiplier and vcore to save power. Remember, It doesn't touch FSB speeds.

The actual CPU speed is calculated as FSB*multiplier. My C2D e7300 has a multiplier of 10 which make it a 2,66GHz CPU at default 266MHz FSB. The lowest powersave speed is reported as 1,6GHz. This results in a multiplier of 6. I didn't measure the actual voltage drops that speedstep produces, but those are not important now. As you know on most CPUs the multiplier is locked, so the max value I can set is 10. The only way for me to overclock is to raise the FSB speed. For example 350MHz FSB makes a 3,5GHz CPU.

Now to the reported speeds in /proc/cpufreq: having set a 350MHz FSB, the reported CPU speed is corretly 3,5GHz. As soon as I load acpi-cpufreq, the reported speeds (cpufreq-info) show the same values as without overclock. Thas is 2,66 MAX and 1,6 MIN. As discussed above, this might be a BIOS bug... I did some tests to confirm that the CPU is still overclocked and what is more important, that scaling is still working. Using hardinfo from community I've discovered, that the CPU really is overclocked, because there is no performance difference before and after loading acpi-cpufreq. I have also established that scaling still works, because setting the powersave governor made the performance drop.

My last point are the BIOS options that disable speedstep on my Asus board. I have options to manually set the multiplier and vcore among others. Once again, what speedstep does is lower those two setting to save power. It is pretty logical that manually forcing any of the two options disables speedstep (and acpi-cpufreq fail to load with a device not found message). Vcore and multiplier settings have to be set at AUTO. This is the case with my Asus board. I have no experience with other boards and brands in this respect.

A higher vcore setting is sometimes required to achieve higher overclocks withou stability issues. But as soon as you raise your vcore you are sacrificing powersave modes. It is save to say that a stable overclock speed that is possible without touching vcore in combination with speedstep is a good option, as your CPU will not consume significantly more power and you still have the nice performance boost. Another good option would be to try and set vcore manually as low as possible. My board goes down to 1,1v (from 1,3 standard for my e7300). I have found out that my CPU is absolutely stable at 1,1v and the default 2,66GHz speed. It is thus possible that actually NOT running speedstep with a fixed very low vcore and stock speed might save you more power than running everything at default. I have not yet tried how far the CPU would overclock with 1,1v. It would be very interesting to see some real world results at the plug. Unfortunately I don't have a wattmeter for this purpose.

So the outcome of this is that the CPU is actually overclocked, only the reported speeds are wrong and that it is pretty easy to calculate how fast the CPU is going for all governors after overclock.

Arch Linux

#1 2008-07-30 21:35:05

Cpufreq and overclocking (bug)

#2 2008-07-30 22:39:08

Re: Cpufreq and overclocking (bug)

#3 2008-07-30 23:03:42

Re: Cpufreq and overclocking (bug)

#4 2008-07-31 00:12:06

Re: Cpufreq and overclocking (bug)

#5 2008-07-31 01:45:24

Re: Cpufreq and overclocking (bug)

#6 2008-08-01 23:07:35

Re: Cpufreq and overclocking (bug)

#7 2008-08-02 20:30:25

Re: Cpufreq and overclocking (bug)

#8 2008-08-03 21:43:56

Re: Cpufreq and overclocking (bug)

#9 2008-08-03 22:05:54

Re: Cpufreq and overclocking (bug)

#10 2008-08-03 22:07:42

Re: Cpufreq and overclocking (bug)

#11 2008-08-03 22:09:40

Re: Cpufreq and overclocking (bug)

#12 2008-08-03 22:18:10

Re: Cpufreq and overclocking (bug)

#13 2008-08-03 23:38:23

Re: Cpufreq and overclocking (bug)

#14 2009-10-17 14:07:36

Re: Cpufreq and overclocking (bug)

#15 2009-10-17 14:52:44

Re: Cpufreq and overclocking (bug)

#16 2010-05-09 08:42:39

Re: Cpufreq and overclocking (bug)

#17 2010-09-27 09:51:44

Re: Cpufreq and overclocking (bug)

Board footer