Need feedback on my undervolting approach

[QuadraticAmoeba]

Hi all, I need some feedback / help on undervolting my Phenom II 955.
I will describe the methodology that I am following below, please let me know if I am on the right path / what I can expect / what I can do better.

1. System configuration is in my signature

2. Disabled Cool and Quiet in the BIOS - otherwise can't test the lower p-states. Will enable it back once I find the lowest voltages for all the states.

3. Downloaded HWMonitor 1.15, IntelTestBurn V2.3 and K10Stat 0.91 and opened all of them

4. Set the processor to P0 state (3200 MHz, 1.35V default)

- Using K10Stat, reduced the CPU voltage by one step (12.5 mV, I think)

- ran IntelTestBurn on Standard settings for 5 iterations (this is the default). Just for the P0, I also ran High for 10 iterations.

- max CPU temperature was 49 C, idle is 32-35C, ambient is around 20 C

- repeat

I could get it down to 1.2125V. At 1.2V, programs started to crash, and I saw display corruption. Interestingly, this screwed up something, and from the next boot, the CPU temperature was showing as 0 C! I had to reflash the BIOS to get back proper CPU temperature.

So now I think that 1.225 V (one step more) is the lowest I can go. I have also run some games at this voltage and it seems stable.

5. Set the processor to P3 state (800 MHz, 0.975V default)

- Using K10stat, force the CPU to 800 MHz, now reduce the voltage by one step

- I am currently running this (0.85V was stable, will test below)

My questions are:

1. Is this a good / optimal methodology?

2. I assume I have to repeat this for the other two p-states as well (I think I will set the frequencies to 1600 MHz and 2400 MHz) - any advice here?

3. How do I permanently set these up, once I have the final settings?

4. The NB voltage is set at 1.1V. I have not attempted to reduce this, is there any advantage to try this voltage?

Thanks for your inputs.

[RedAE102]

Being an avid undervolter, I'll bite. Unfortunately, I can't help with k10stat since I don't have a K10 and have never directly dealt with one. Your methods seem sound, except I'd recommend using Prime95 or Orthos (which is Prime95 with a GUI tailored to stability testing), which do a good job of stressing the RAM as well as the cache. Orthos will halt the stress test as soon as it encounters instability (a rounding error greater than expected). Considering I can only get my aging 65nm Brisbane G2 down to 1.225V at 2.8 GHz, I would think you should be able to go much lower than 0.850V @ 800MHz, considering the Brisbane runs 1.2 GHz @ 0.775 V, the lowest VID available.

Hopefully your fellow K10 owners can help with getting everything set up in k10stat, as well as comment on NB voltage.

A good place to start with finding voltages for 1.6 and 2.4 GHz is to graph voltage versus speed for 800MHz and 3.2 GHz and add a linear trendline between them. Where the voltage line hits 1.6 and 2.4, start with those voltages, and you can guarantee you won't be starting with too low a voltage.

[mrle]

1. I am also using K10stat, and the methods you use are fine, I used a very similar approach.

2. Same as for P0 and P3.

3. You cannot permanently set it, every time you restart, these settings are lost. But you can save a "profile" in K10stat and then create a shortcut in Windows Startup folder, like this: "C:\AMD\K10stat\K10STAT.exe -loadprofile:1 -nowindow -loadpstate". This will silently load your custom P-states on Windows boot.

4. Don't know if there's any benefit in lowering it. On my AM2 board NB voltage is tied to CPU voltage, so it goes down to 0.925V.

edit: I don't think you need to disable C'n'Q in BIOS. If you want to test one particular P-state, you can use K10stat to force it, and when you're done, just disable K10stat's Clock Control function (set it to Auto). That way you'll still be using C'n'Q but with your custom multipliers and voltages.

[QuadraticAmoeba]

Thank you all for your informative replies!

- I will try the K10stat startup and also try without disabling C'n'Q in BIOS

- The linear trendline approach makes a lot of sense, it also appeals to the geek in me

- I chose IntelTestBurn simply because it is much faster - 15 minutes of testing compared to hours with Orthos. But I take your feedback and will run Orthos to ensure that my voltages are high enough

- On my board, you can set NB and CPU voltage separately, but I'm not going to touch NB until I have the CPU all setup

I will post back here with my results in a few days. Thanks once again.

[RedAE102]

- I chose IntelTestBurn simply because it is much faster - 15 minutes of testing compared to hours with Orthos. But I take your feedback and will run Orthos to ensure that my voltages are high enough

In my own experience, 10 minutes of Orthos has usually been enough to uncover instabilities. The anomaly is my retired Brisbane G1 Athlon 64 X2 3600+, which would get errors after 8+ hours. In every other processor I've used it on, including my Brisbane G2 5400 (2.8 GHz), any instability turned up in the first 10 minutes. If it made it through the first 10 minutes, it would run for 24+ hours with no errors.

[QuadraticAmoeba]

Thanks to all the folks who helped me. I wasn't able to spend much time on this, but here is my current status:

Freq Voltage
3200 1.225
2400 1.075
1600 0.925
800 0.775

I am using the linear MHz/mV assumption, and it is working out beautifully!

I have still not figured out how to get these set at startup - I put a shortcut to K10Stat in the Startup folder, but I don't see it working..

[QuadraticAmoeba]

Apologies for replying to my own post. I thought I'll try this one more time before I head off to bed.
And it worked!

Here is what I got:

Freq Voltage
3200 1.225
2400 1.0625
1600 0.900
800 0.7375

So that's more than 25% voltage drop at 800 MHz. I did some very fast power calculations, and I am down at least 8% at full load, and possibly more at idle.

I also figured out the K10stat - I just added "-StayOnTray" to the startup command, and it worked!

Thanks once again to all the fine folk on this forum.