Undervolting with kernel 2.6 [Done!]

[Fran]

I used to have my athlon 64 undervolted to 1.4V with kernel 2.6.8, which was nice because with the cpu fan at 1400rpm (zero noise) the max temp was 53ºC at full load. With that kernel, when I switched the cpu governor to powersave and then back to performance, the undervolting was gone and the cpu went to default 1.5V, but at least I could have my cpu at 1.4 when I booted (or if I didn't change the governor).

But since 2.6.9, the cpu voltage is ALWAYS set at 1.5V when the frequency is 2.2GHz, no matter what I select in the bios. That forces me to increase the fan speed up to 2500rpm everytime I compile something , if I don't want the temp to exceed 63ºC .

I tried setting the default governor to userspace and performance, but that didn't help. I also tried removing the "ondemand" governor (the only visible difference between 2.6.8 and 2.6.9 in cpufreq), but it didn't help either.

Is there any way to choose the voltage of the CPU when it is put at max freq?

Thanks.

(edit) I've been investigating and I suposse that I can set manually the voltage table patching powernow-k8.c... but... it's scary .

(edit again) Done! . I modified powernow-k8.c and now my frequencies and voltages are:

2.2GHz ---> 1.4V
1.8GHz ---> 1.3V
1.0GHz ---> 1.0V

Even with the "ondemand" governor

[brihall]

Could you post (or give a link to) your modified powernow-k8.c? A patch would be nice. Even nicer, would be a way to set these voltage values via a /sys interface!

[Fran]

My modified version is here ;). I only added one line, changing the value of data->powernow_table[0].index, which is read from the bios, to the value I'm interested in. The voltage set (in mV) by the kernel for each frequency is calculated as 1550-(data->powernow_table[j].index >> 8)*25. The default value read from the bios for 2.2GHz is (2 << 8), so we obtain 1550-2*25=1500. In my case, I want 1.4V for the highest freq. By adding (4 << 8) to that value we obtain 1550-6*25=1400mV.

If you want to try with lower voltages (1.375 or 1.35, for example), you can respectively use (5 << 8) or (6 << 8) instead of (4 << 8) ;).

Of course you can also change the voltages for 1.8GHz (data->powernow_table[1].index) and for 1GHz (data->powernow_table[2].index) if you want.

And of course you can find a better place for changing the value (for example, in the place where the values are read from the bios) or just change the value for (6 << 8) instead of adding (4 << 8)... but this is the first thing I tried, it worked, and you know... if it works, don't touch it :P.

The patch (edit: for 2.6.11):

Code: Select all

*** powernow-k8.c       2005-03-05 16:21:02.000000000 +0100
--- powernow-k8.c       2005-03-05 16:44:54.000000000 +0100
***************
*** 522,527 ****
--- 522,529 ----
  static void print_basics(struct powernow_k8_data *data)
  {
        int j;
+       if (data->numps >= 0)
+               data->powernow_table[0].index = data->powernow_table[0].index + (4 << 8);
        for (j = 0; j < data->numps; j++) {
                if (data->powernow_table[j].frequency != CPUFREQ_ENTRY_INVALID)
                        printk(KERN_INFO PFX "   %d : fid 0x%x (%d MHz), vid 0x%x (%d mV)\n", j,

I suppose I don't need to say it but: use this UNDER YOUR OWN RESPONSABILITY :). Anyway, there should be no danger if you understand what you're doing. I've been using it since yesterday and works great :D.

(edited to correct a little thing)

[brihall]

OK, here's my version of the patch. After browsing the 'net for "undervoltage testimonials", I determined that the largest, safest undervolt was the following set (for a CG 3000+ 2GHz .13 micron chip):

2000 MHz 1.300 V
1800 MHz 1.200 V
1000 MHz 0.800 V

I have been running with these values for the past hour, emerging 11 packages while running foldingathome with asm options enabled and make -j3 kernel compiles. No problems whatsoever.

I did run the CPU fanless while doing these loading tests, but the max temp of 63C seemed too high to me. With the stock CPU HSF @ 5V I max out at 40C! Mid-tower steel case, 1 80mm PSU + 1 120mm exhaust.

The patch:

Code: Select all

--- /usr/src/linux/arch/i386/kernel/cpu/cpufreq/powernow-k8.c   2004-12-04 17:55:37.000000000 -0700
+++ powernow-k8-bwh.c   2005-01-21 18:21:39.741467104 -0700
@@ -516,7 +516,10 @@

 static void print_basics(struct powernow_k8_data *data)
 {
-       int j;
+        data->powernow_table[0].index =        data->powernow_table[0].index | 0x1c00; // 0.800V
+       data->powernow_table[1].index = data->powernow_table[1].index | 0x0e00; // 1.200V
+       data->powernow_table[2].index = data->powernow_table[2].index | 0x0a00; // 1.300V
+        int j;
        for (j = 0; j < data->numps; j++) {
                if (data->powernow_table[j].frequency != CPUFREQ_ENTRY_INVALID)
                        printk(KERN_INFO PFX "   %d : fid 0x%x (%d MHz), vid 0x%x (%d mV)\n", j,
@@ -577,6 +580,7 @@
        dprintk("cfid 0x%x, cvid 0x%x\n", data->currfid, data->currvid);
        data->powernow_table = powernow_table;
        print_basics(data);
+       return 0; /* allow override of VID values in print_basics() */

        for (j = 0; j < data->numps; j++)
                if ((pst[j].fid==data->currfid) && (pst[j].vid==data->currvid))

[tnt]

+ data->powernow_table[0].index = data->powernow_table[0].index | 0x1c00; // 0.800V

I guess this is 0.850V...

[brihall]

Yes, it was. Here's a slightly updated version of the patch. I had a lockup while idle that I attributed to the lower voltage, so I've raised the minimum to 0.90V. However, I have also been able to lower the voltage required for max speed stable operation. For standard desktop usage patterns, the idle voltage is important for overall power consumption, and the max speed voltage has more of an impact on temperature.

(2.6.11-ck2 kernel, 24x7)

cat /sys/devices/system/cpu/cpu0/cpufreq/stats/time_in_state
1000000 488407955
1800000 10515447
2000000 14432428


--- arch/i386/kernel/cpu/cpufreq/powernow-k8.c.orig 2005-03-02 00:38:10.000000000 -0700
+++ arch/i386/kernel/cpu/cpufreq/powernow-k8.c 2005-03-06 16:07:55.000000000 -0700
@@ -521,7 +518,13 @@
static void print_basics(struct powernow_k8_data *data)
{
- int j;
+ data->powernow_table[0].index = data->powernow_table[0].index = 0x1a02; // 0.900V
+ data->powernow_table[1].index= data->powernow_table[1].index = 0x120a; // 1.100V
+ data->powernow_table[2].index= data->powernow_table[2].index = 0x0e0c; // 1.200V
+ int j;
for (j = 0; j < data->numps; j++) {
if (data->powernow_table[j].frequency != CPUFREQ_ENTRY_INVALID)
printk(KERN_INFO PFX " %d : fid 0x%x (%d MHz), vid 0x%x (%d mV)\n", j,
@@ -582,6 +585,7 @@
dprintk("cfid 0x%x, cvid 0x%x\n", data->currfid, data->currvid);
data->powernow_table = powernow_table;
print_basics(data);
+ return 0; /* allow override of VID values in print_basics() */

for (j = 0; j < data->numps; j++)
if ((pst[j].fid==data->currfid) && (pst[j].vid==data->currvid))

[ikaro]

No patch needed.

o Go to the bios and set the default cpu volt to 1.3.
o use the ondemand governor

[Fran]

No patch needed.

o Go to the bios and set the default cpu volt to 1.3.
o use the ondemand governor

Maybe with your mobo and bios. Not with mine (MSI K8T NEO FSR with bios 1.7), since 2.6.9. I've always had 1.4V as the default voltage in the bios, and if I forget the patch when I compile a new kernel, the voltage is set to 1.5.

BTW, this thread is from october, i'll edit it with the patch for 2.6.11 .

[ikaro]

thats because the bios shows the wrong voltages.
so 1.5 its actually 1.4.. this is something I read at extremeOverclocking.com(? not sure about the name )
it just prevents you from overclocking too much, because you can still reach decent speeds if the cooling is good.

[Fran]

Mmm... I don't understand very well what you mean... but I have an external tempreature sensor and I do notice the huge difference in temperatures (almost 10º at full load) when it's at 1.5V or at 1.4V. The sensor can't lie. And if I set the voltage in the bios at 1.4 and boot windows, the voltage is kept at 1.4, the temperature sensor confirms it. So it's a kernel problem.

[tnt]

No patch needed.

o Go to the bios and set the default cpu volt to 1.3.
o use the ondemand governor

BIOS settings do not affect C'n'Q (powernow-k8) because multipliers and voltages for it are read from table later stored in:

data->powernow_table[x].index

and they override values manualy set in BIOS.
So, the only way to instruct kernel to set some other voltage then default (for that CPU state) is to modify values in that table.

[Jtb]

why do you change the voltage in static void print_basics?
Wouldn't it much better if you change them directly after reading them?
I change my voltage in powernow_k8_cpu_init_acpi and fill_powernow_table directly before "powernow_table.xy ="

[Fran]

why do you change the voltage in static void print_basics?
Wouldn't it much better if you change them directly after reading them?
I change my voltage in powernow_k8_cpu_init_acpi and fill_powernow_table directly before "powernow_table.xy ="

When I made the patch the first time I didn't want to search through the whole code of powernow-k8.c to find a good place to change the voltages. So I put it in the first place I found that I knew it would pass at least once, and only once. As I said above:

And of course you can find a better place for changing the value (for example, in the place where the values are read from the bios) [...]... but this is the first thing I tried, it worked, and you know... if it works, don't touch it :P.

;).

[Jtb]

What about creating a undervoltage-parameter for the powernow-k8-modul?
I'm no kernel-programmer but I seems simply to realize..

[servo888]

How about the 2.6.15 kernel? Everything looks different...

[tnt]

I've made that hack last time on the 2.6.13, but as far as I can see there is no difference with 2.6.15 (just line numbering is somewhat shifted):

Code: Select all

diff powernow-k8.c.2.6.13 powernow-k8.c.2.6.13.original


537,538d536
<       data->powernow_table[0].index = data->powernow_table[0].index + (5 << 8); // 1.275V
<       data->powernow_table[1].index = data->powernow_table[1].index + (4 << 8); // 1.000V
599,601d596
<
<
<       return 0; /* allow override of VID values in print_basics() */

[Fran]

How about the 2.6.15 kernel? Everything looks different...

Yeah, powernow-k8.c changes a little with every kernel version. But the lines you must add are the same. Just look for the proper location (print_basics or wherever you want to put them).

PS: This thread never dies

[tnt]

This thread never dies

because it's so useful

[Evangelion]

few questions:

a) why doesn't the powernow-module scale the voltage (like it should)?

b) where do you check the voltage of the CPU in a running system?

[quax]

Code: Select all

*** powernow-k8.c       2005-03-05 16:21:02.000000000 +0100
--- powernow-k8.c       2005-03-05 16:44:54.000000000 +0100
***************
*** 522,527 ****
--- 522,529 ----
  static void print_basics(struct powernow_k8_data *data)
  {
        int j;
+       if (data->numps >= 0)
+               data->powernow_table[0].index = data->powernow_table[0].index + (4 << 8);
        for (j = 0; j < data->numps; j++) {
                if (data->powernow_table[j].frequency != CPUFREQ_ENTRY_INVALID)
                        printk(KERN_INFO PFX "   %d : fid 0x%x (%d MHz), vid 0x%x (%d mV)\n", j,

I think, better is to edit DSDT an overide old table with new via CONFIG_ACPI_CUSTOM_DSDT kernel config option.

You need add or modify _PSS method and others related methods (_PPC). Nice example is in the ACPI Specification v3.0 on pages 274 - 277.

[Fran]

a) why doesn't the powernow-module scale the voltage (like it should)?

powernow-k8 does scale the voltage, following a table that it reads from the bios. But if you want lower voltages than those in that table, you must patch the kernel.

b) where do you check the voltage of the CPU in a running system?

You can try with lm-sensors, or simply look at /sys/bus/i2c/devices/*/in*_input. One of those files contains Vcore.

I think, better is to edit DSDT an overide old table with new via CONFIG_ACPI_CUSTOM_DSDT kernel config option.

You need add or modify _PSS method and others related methods (_PPC). Nice example is in the ACPI Specification v3.0 on pages 274 - 277.

Probably a more elegant solution, but it seems way more complicated (at least looking at the pdf, gentoo wiki is down right now).

[quax]

powernow-k8 does scale the voltage, following a table that it reads from the bios. But if you want lower voltages than those in that table, you must patch the kernel.

Reading PST table from bios is deprecated, ACPI way is prefered. PST table supports only one CPU. powernow-k8 module read first APCI.

Probably a more elegant solution, but it seems way more complicated (at least looking at the pdf, gentoo wiki is down right now).

My laptop have a buggy DSDT table (Acer Aspire 1511LMi) and I was spend too many hours with repairing this table and patching kernel with APCI from sf.net. Adding _PSS method isn't problem for me, but i don't know right values for the TransitionLatency, BusMasterLatency and Control parameters.

[tnt]

cpufreq at lists dot linux dot org dot uk

[morfic]

please try the following patch if you would like the ability to change the voltage at boot time or whenever inserting powernow-k8 module

patch kernel 2.6.16-rc2 (only tested on vanilla) with this patch: http://dev.gentoo.org/~morfic/powernow- ... 6-rc2.diff

add powernow-k8.vcore_list=value1,value2,value3 to the kernel command line in grub or lilo

or build powernow-k8 as module and do 'modprobe powernow-k8 vcore_list=value1,value2,value3

please only specify any new voltages for each powernow-k8 step you have

dmesg | grep powernow will show current values/number of states you have

value is in millivolt, so to change from 1.3 to 1.1 you would specify 1100

i will now paste my dmesg to illustrate how this works on a turion64 with 2 states

morfic@localhost64 ~ $ dmesg | grep powernow
Bootdata ok (command line is root=/dev/hda7 vga=792 powernow-k8.vcore_list=1100,900)
Kernel command line: root=/dev/hda7 vga=792 powernow-k8.vcore_list=1100,900
powernow-k8: Found 1 AMD Athlon 64 / Opteron processors (version 1.60.0)
powernow-k8: Request Voltage for id0 is 1100 mV default voltage is 1450
powernow-k8: Performing required adjustment of 350 mV or 14 25mV steps
powernow-k8: Request Voltage for id1 is 900 mV default voltage is 1000
powernow-k8: Performing required adjustment of 100 mV or 4 25mV steps
powernow-k8: 0 : fid 0x8 (1600 MHz), vid 0x12 (1100 mV)
powernow-k8: 1 : fid 0x0 (800 MHz), vid 0x1a (900 mV)
powernow-k8: ph2 null fid transition 0x8
morfic@localhost64 ~ $

the ph2 line means we trigger a transition where the frequency has not changed

i have made the patch because calculating it was too cumbersome fr me and friends, aside my default voltage is 1.45V in linux while i used the spec of 1.5V so i was always off by .05V, the patch will adjust the voltage by the difference of currently read from bios from requested voltage, much more failsafe i think

as usual, neither i nor gentoo grant no guarantees and warrantees, use fhis patch is entirely at your own risk, improper use of this patch can lead to system instabilities and damage your hardware, by downloading the and applying the patch you acknowledge this.

i would appreciate user feedback, especially from people with 3 or more power states

note: the power usage in /proc/acpi/processor/cpu*/power is hardcoded and will not change when you lower your vcore
note: if you build powernow-k8 as module you will not automatically use your usual scaling governor, modprobe powernow-k8 and set your usual scaling governor

[morfic]

forgot to add, 1.45V@1600Mhz means 90C compiles
while 1.1V@1600MHz mean 60C compiles

hot little turion, i know