Updated: Tuning Cool'n'Quiet: Maximize Power And Performance
-
Page 1:Introduction
-
Page 2:Fine Tuning Power Management
-
Page 3:Test System
-
Page 4:Update: A “Broken” Cool'n'Quiet Implementation
-
Page 5:Assessing Voltage
-
Page 6:Tweaking And Undervolting
-
Page 7:Voltage Ramps, Continued
-
Page 8: Measured Power Consumption And Methodology
-
Page 9:Test Results: Idle
-
Page 10:Test Results: Idle, Continued
-
Page 11:Test Results: Load
-
Page 12:Voltage Ramp And Power Consumption
-
Page 13:Voltage Ramp And Power Consumption, Continued
-
Page 14:Phenom X4 955 And Conclusion
Update: A “Broken” Cool'n'Quiet Implementation
Update: We've received feedback from AMD regarding the power numbers and resulting analysis generated using this piece's platform. We are currently looking at other hardware combination in order to either validate or correct the initial conclusions drawn.
Update 2: After months of retesting, we finally chased down the problem to a BIOS bug with the TA790GX 128M that caused incorrect P-state settings to be applied. Updating the board's BIOS to the latest beta version fixed this problem. So, Cool'n'Quiet is not at fault for the results we saw previously. Instead, what we saw was a broken Cool'n'Quiet implementation. This does affect power consumption measurements with default settings for the Athlon X2 7750. But the other results are still valid, since they were taken with different motherboards that do not exhibit this behavior. The article have been edited to reflect this.
There are very subtle, yet very significant differences between these processors' implementations of Cool'n'Quiet technology. Let's check their respective default voltages, according to CPU-Z:
Default Clock | Default Voltage | CnQ Clock | CnQ Voltage | |
---|---|---|---|---|
Athlon | 2.7 GHz | 1.312V | 1.35 GHz | 1.36V |
Athlon | 2.8 GHz | 1.280V | 1.4 GHz | 1.04V |
Athlon | 3 GHz | 1.408V | 800 MHz | 1.056V |
Phenom | 2.6 GHz | 1.28V | 800 MHz | 1.184V |
Phenom | 3 GHz | 1.424V | 800 MHz | 1.072V |
Phenom | 3.2 GHz | 1.424V | 800 MHz | 1.04V |
Using the TA790GX 128M, the Athlon X2 7750 default voltages set for higher p-states (idle mode) are actually higher than the default voltage for lower p-states (performance mode). Let's see how that strange anomaly affects power consumption at idle. On a motherboard with a proper Cool'n'Quiet implementation, the Athlon X2 7750 drops to 1.04V at idle, which is the same as the Athlon X2 7850.
Default System Power Consumption | CnQ-High Performance | CnQ-Balanced | |
---|---|---|---|
Athlon X2 7750 | 81 watts | 73 watts | 79 watts |
After the BIOS fix was applied, a quick test shows power consumption is about the same (74 watts average). In that regard, power consumption with and without Cool'n'Quiet is not that much different for the Athlon X2 7750 on a Biostar TA790GX 128M.
Athlon X2, Phenom X4, and Athlon II processors will use different clock rates for different cores, independent of each other. Though technically interesting, this adversely affects single-thread, single-core performance when Cool'n'Quiet is enabled, particularly in Windows Vista. Windows 7's scheduler manages threads more efficiently (bounding them around from core to core less often), based on our testing of Intel CPUs under the new operating system, so we'd expect older AMD CPUs to suffer less after an upgrade to 7.
Disclaimer note: of course, changing p-states settings is not without risks. An overly aggressive configuration will likely cause instability/crashes. There's also no guarantee that the savings will be worth your effort. In this regard, hacking p-state settings is no different than overclocking in that you are running the processor outside of its manufacturer's recommended specification. Continue at your own risk.
If you are unsatisfied with the default power management settings, there's a way to squeeze more power savings. As noted before, applications like K10Stat and PhenomMSRTweaker allow you to change this behavior so that clock changes are applied to all cores. You can also use them to edit register values used by Cool'n'Quiet, overriding the default values.
How about newer processors? Although the Phenom II X4 and X3 CPUs have changed the behavior of Cool'n'Quiet (proper setting of voltages and synchronous clock changes for all cores), you can still use those two applications to fine tune voltage settings.
There are other advantages associated with manipulating p-states settings. Unlike tweaking BIOS voltage settings, you can directly monitor the processor's voltage in real time. That means vdroop and offset are already taken into account. In addition, these settings are only applied when you want them to be applied. They are not saved in the BIOS. If a setting is too aggressive and causes problems (the familiar BSOD), a reset will revert their values to default.
- Introduction
- Fine Tuning Power Management
- Test System
- Update: A “Broken” Cool'n'Quiet Implementation
- Assessing Voltage
- Tweaking And Undervolting
- Voltage Ramps, Continued
- Measured Power Consumption And Methodology
- Test Results: Idle
- Test Results: Idle, Continued
- Test Results: Load
- Voltage Ramp And Power Consumption
- Voltage Ramp And Power Consumption, Continued
- Phenom X4 955 And Conclusion
All Cool'n'Quiet 2.0 processors should work with K10Stat, so you should be able to use K10Stat with the Energy Efficient series.
For those unfamiliar with K10Stat, here are some additional switches/options (insert these via the properties menu of your K10Stat shortcut).
-lp:# - load profile# and write to MSR (activates a certain profile).
-nw - Start K10stat with NoWindow (don't display window).
-ClkCtrl:#(0-4) - Enable Clock Control (activates clock/core control).
0:No Control
1:UnGanged.
2:Ganged (Load of Highest core).
3:Ganged (Average load of all cores).
4:Ganged (Load of Lowest core).
-StayOnTray - puts K10Stat on the system tray. Use this option with “-nw” to keep K10Stat running. Very useful if you use sleep/hibernate.
To those interested in Core i7/Core i5, a similar article is in the pipeline. We wanted to include the 32 nm refresh too, in addition to Core2Duo and Core2Quads, so we have to wait until (final) samples are out.
About performance, performance differences and actual power consumption will be in the 2nd part - you'll see it soon. You might be (pleasantly) surprised.
Thanks. Off to some more GPGPU testing.
People that say that tend to be fairly annoying.. Anyways I finally found a mediafire link.. all the geocities and rapid share links are dead..
http://www.mediafire.com/?dywh3zhmk41