[ADVANCED] RMClock Powersaving, Whining Stop

[7oby]

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:
The "Undervolting" Guide

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...tecture_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/det...p?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/a...s/eng/3726.htm

Originally Posted by intel

For time-bound applications, where we have a fixed amount of data to process and the application is well multi-threaded using full CPU resources, it is better to finish faster and then go into idle in order to save power. Of course, the recommendation is to allow Intel® SpeedStep™ technology and the OS to determine the ideal CPU frequency for your application.

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.

[rubenvb]

Nice info here!

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

[royk50]

good reading, 10x alot

[paper_wastage]

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

[Kreeeee]

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

[7oby]

Originally Posted by Kreeeee

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.

[flipfire]

Great research you have there

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

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:

get that work done as fast as possible and jump back to low power C4

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.

[7oby]

Originally Posted by flipfire

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.

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...ol10_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/documentati...n-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/a...s/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/a...s/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.

Originally Posted by flipfire

SuperLFM multiplier should only be used for Idling, NOT FOR CPU LOAD.

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.

Originally Posted by flipfire

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.

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

Originally Posted by flipfire

Try driving a car with only brakes and full throttle
......... actually that would be fun.

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.

[Kreeeee]

Originally Posted by 7oby

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

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.

[7oby]

Originally Posted by Kreeeee

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