Centrino: undervolting and further reducing heat

Centrino: undervolting and further reducing heat

[Daniel Bonniot]

Hi,

I have a dothan 1.6 GHz processor, and i've been investigating how to reduce 
heat to minimize fan usage. One very promising avenue is undervolting. 
According to several online articles, there is a large potential for 
undervolting. I ended up modifying speedstep-centrino.c to test it out. 
Results are very positive: it can run under heavy load at 1.6Ghz with only 
972mV (down from 1340mV). This decreased temperature by 10°C! At 600MHz, the 
minimum 700mV was sufficient.

I understand that the driver needs to be conservative by default, but given 
the huge benefits, it would make a lot of sense to offer a way to specify 
lower voltages to those who want to. Has anybody investigated this 
possibility, or already started working on it? What would be a good interface 
to specify the voltages?


Independently of this feature, I wonder how it's possible to further reduce 
consumption when the system is mostly ideal, and even 600MHz is more than 
needed. The idea being to let the system cool down even more in that case. Is 
there any existing way to achive this on linux? One possibility I found in the 
intel docs is the IA32_THERM_CONTROL MSR, which can reduce clock speed by 
12.5% to 87.5%. Am I right in thinking that it could be combined with 
speedstep to achieve even lower power consumption states? Are there other 
possibilities?

Cheers,

Daniel

Re: Centrino: undervolting and further reducing heat

[Nebojsa Trpkovic]

I use undervolted AMD64 on two gentoo boxes for months:

1. socket 754 nforce3-250gb, athlon64 3000+ (Newcastle core) overclocked
to 2.4GHz and undervolted to run @1.45V

2. socket 939 nforce4 ultra,  athlon64 3000+ (Winchester core)
undervolted to run @1V at 1GHz and @1.275V at 1.8GHz (stable voltages
suficient to run prime95 for hours were 0.9V at 1GHz and 1.175V at
1.8GHz but I like to be 110% safe so I've left extra 0.1V just in case)

Ofcorse, I had to manualy modify powernow-k8.c

;)




Daniel Bonniot wrote:

>
> Hi,
>
> I have a dothan 1.6 GHz processor, and i've been investigating how to
> reduce heat to minimize fan usage. One very promising avenue is
> undervolting. According to several online articles, there is a large
> potential for undervolting. I ended up modifying speedstep-centrino.c
> to test it out. Results are very positive: it can run under heavy load
> at 1.6Ghz with only 972mV (down from 1340mV). This decreased
> temperature by 10���C! At 600MHz, the minimum 700mV was sufficient.
>
> I understand that the driver needs to be conservative by default, but
> given the huge benefits, it would make a lot of sense to offer a way
> to specify lower voltages to those who want to. Has anybody
> investigated this possibility, or already started working on it? What
> would be a good interface to specify the voltages?
>
>
> Independently of this feature, I wonder how it's possible to further
> reduce consumption when the system is mostly ideal, and even 600MHz is
> more than needed. The idea being to let the system cool down even more
> in that case. Is there any existing way to achive this on linux? One
> possibility I found in the intel docs is the IA32_THERM_CONTROL MSR,
> which can reduce clock speed by 12.5% to 87.5%. Am I right in thinking
> that it could be combined with speedstep to achieve even lower power
> consumption states? Are there other possibilities?
>
> Cheers,
>
> Daniel
>
>
> _______________________________________________
> Cpufreq mailing list
> Cpufreq@lists.linux.org.uk
> http://lists.linux.org.uk/mailman/listinfo/cpufreq
>

Re: Centrino: undervolting and further reducing heat

[Aaron Spettl]

Hi

I've got a banias 1.6 GHz processor and like Daniel I've reduced the
voltages. My multiplicators and voltages:
  6x @ 0.780V (default: 0.988V)
10x @ 0.956V
12x @ 1.004V
14x @ 1.116V
16x @ 1.244V (default: 1.484V)
So far I've found nobody who needs higher settings on a banias cpu (I'm
admin on a german notebook forum and have seen many values).

Dothan cpus need less power, in my experience (very) safe values are:
  6x @ 0.732V
10x @ 0.860V
12x @ 0.988V
14x @ 1.052V
16x @ 1.212V
17x (?)
18x (?)
As you can see with Daniels cpu (6x @ 0.700V, 16x @ 0.972) these
voltages should be no problem for anyone, so I suppose it would be safe
to set the default values to these.
I think my banias values should be increased by one step (0.016V) to be
sure.

Idea:
Is it possible to set a kernel option like "reduce cpu voltage", maybe
even an "aggressive" option? We could leave the original tables for
default and activate the lower voltages only if this option is set. A
kernel boot parameter could override this option for one session, so
people could test it easily.

Regards
Aaron


Nebojsa Trpkovic wrote:

>I use undervolted AMD64 on two gentoo boxes for months:
>
>1. socket 754 nforce3-250gb, athlon64 3000+ (Newcastle core) overclocked
>to 2.4GHz and undervolted to run @1.45V
>
>2. socket 939 nforce4 ultra,  athlon64 3000+ (Winchester core)
>undervolted to run @1V at 1GHz and @1.275V at 1.8GHz (stable voltages
>suficient to run prime95 for hours were 0.9V at 1GHz and 1.175V at
>1.8GHz but I like to be 110% safe so I've left extra 0.1V just in case)
>
>Ofcorse, I had to manualy modify powernow-k8.c
>
>;)
>
>
>
>
>Daniel Bonniot wrote:
>
>  
>
>>Hi,
>>
>>I have a dothan 1.6 GHz processor, and i've been investigating how to
>>reduce heat to minimize fan usage. One very promising avenue is
>>undervolting. According to several online articles, there is a large
>>potential for undervolting. I ended up modifying speedstep-centrino.c
>>to test it out. Results are very positive: it can run under heavy load
>>at 1.6Ghz with only 972mV (down from 1340mV). This decreased
>>temperature by 10°C! At 600MHz, the minimum 700mV was sufficient.
>>
>>I understand that the driver needs to be conservative by default, but
>>given the huge benefits, it would make a lot of sense to offer a way
>>to specify lower voltages to those who want to. Has anybody
>>investigated this possibility, or already started working on it? What
>>would be a good interface to specify the voltages?
>>
>>
>>Independently of this feature, I wonder how it's possible to further
>>reduce consumption when the system is mostly ideal, and even 600MHz is
>>more than needed. The idea being to let the system cool down even more
>>in that case. Is there any existing way to achive this on linux? One
>>possibility I found in the intel docs is the IA32_THERM_CONTROL MSR,
>>which can reduce clock speed by 12.5% to 87.5%. Am I right in thinking
>>that it could be combined with speedstep to achieve even lower power
>>consumption states? Are there other possibilities?
>>
>>Cheers,
>>
>>Daniel
>>
>>
>>_______________________________________________
>>Cpufreq mailing list
>>Cpufreq@lists.linux.org.uk
>>http://lists.linux.org.uk/mailman/listinfo/cpufreq
>>
>>    
>>
>
>_______________________________________________
>Cpufreq mailing list
>Cpufreq@lists.linux.org.uk
>http://lists.linux.org.uk/mailman/listinfo/cpufreq
>
>  
>

Re: Centrino: undervolting and further reducing heat

[Wes Felter]

Daniel Bonniot wrote:

> Independently of this feature, I wonder how it's possible to further 
> reduce consumption when the system is mostly ideal, and even 600MHz is 
> more than needed. The idea being to let the system cool down even more 
> in that case. Is there any existing way to achive this on linux? One 
> possibility I found in the intel docs is the IA32_THERM_CONTROL MSR, 
> which can reduce clock speed by 12.5% to 87.5%. Am I right in thinking 
> that it could be combined with speedstep to achieve even lower power 
> consumption states? Are there other possibilities?

Throttling isn't going to help if the system is already mostly halted. 
You're already doing as much as can be done. (Unless there is an 
undocumented setting even lower than 600MHz...)

Wes Felter - wesley@felter.org

Re: Centrino: undervolting and further reducing heat

[Daniel Bonniot]

> Throttling isn't going to help if the system is already mostly halted.

Yes, I tried and it makes very little difference indeed, if any.

> You're already doing as much as can be done. (Unless there is an 
> undocumented setting even lower than 600MHz...)

I wonder if there is a way to reduce the speed of other components of the 
system (the memory is quite hot, for instance). I'm pretty sure the graphic 
card (ATI) can be underclocked too.

Daniel

Re: Centrino: undervolting and further reducing heat

[Bruno Ducrot]

On Thu, May 19, 2005 at 01:15:13AM +0200, Daniel Bonniot wrote:
> I wonder if there is a way to reduce the speed of other components of the 
> system (the memory is quite hot, for instance). I'm pretty sure the graphic 
> card (ATI) can be underclocked too.

IIRC recent xorg do have some power saving options for radeon.
I don't remember if its in CVS or if patches have been committed though,
or if this work only for mobile versions.

-- 
Bruno Ducrot

--  Which is worse:  ignorance or apathy?
--  Don't know.  Don't care.

Re: Centrino: undervolting and further reducing heat

[Daniel Bonniot]

> IIRC recent xorg do have some power saving options for radeon.
> I don't remember if its in CVS or if patches have been committed though,
> or if this work only for mobile versions.

Yes, I'm using Xorg 6.8.2 (-10 from Ubuntu), and the xorg driver does have 
power saving options, although they are not on by default (it's the 
DynamicClocks option). The proprietary drivers enable it by default. My 
preliminary experiments seem to show that you can save further by 
underclocking the radeon card (a Mobility in my case).

Daniel