Thinking about undervolting a 6-core 960T

[mikeymikec]

According to CPU-Z, the core voltage seems to vary between 1.284v and 1.44v (though even when Turbo Core / Boost / whatever kicks in, bringing the core speed up to 3.4GHz, the core voltage doesn't necessarily increase sometimes, perhaps that's an inaccuracy).

I'm using the ASUS M4A89GTD PRO/USB3 board, no overclocking. According to the manual my BIOS has the following settings wrt core voltage:

CPU & NB Voltage Mode: [Offset]
CPU Offset Voltage: [Auto]
CPU/NB Offset Voltage [Auto]
CPU VDDA Voltage [Auto]
DRAM Voltage [Auto]
HT Voltage [Auto]
NB Voltage [Auto]
I have DDR3-1600 dual-channel RAM, in case that is relevant.

All 6 cores are activated since new for a few months without any problems. How should I start playing with these settings?

I'm not interested in overclocking. I'm just interested in maintaining current performance/stability as efficiently as possible (partly in the hope of reducing waste heat output and saving a little on electricity).

 

[Puppies04]

First make sure you have stress tested your PC to confirm it is 100% stable in its current state then set the CPU offset voltage to -0.01v and stress test again, if stable move to -0.02v and repeat.

 

[mikeymikec]

I wonder why it is that I get a lower core voltage reading after ten minutes of Prime95 than when idle. Bizarre. Admittedly it's not by much, 0.01 and some change, but it's still weird.

(Using CPU-Z 1.59 to monitor the core voltage while in Windows)

 

[jvroig]

I can answer your questions, courtesy of a 1090T I've played around with (overvolting, undervoloting, overclocking, underclocking, etc). All figures I will quote is based on my setup (1090T, MSI 990FX-A GD80). I recorded a lot of my test data on a spreadsheet, so I don't actually need to get the figures from memory. I knew recording them would somehow be of use one day

First:

I wonder why it is that I get a lower core voltage reading after ten minutes of Prime95 than when idle. Bizarre.

At stock settings , the vcore spike to ~1.45V is actually due to turbo. With all cores maxed out, vcore only goes up to about 1.376V, because by then Turbo is certainly not being triggered anymore.

According to CPU-Z, the core voltage seems to vary between 1.284v and 1.44v (though even when Turbo Core / Boost / whatever kicks in, bringing the core speed up to 3.4GHz, the core voltage doesn't necessarily increase sometimes, perhaps that's an inaccuracy).

See answer above. Same thing. The spikes to greater than 1.4V are due to Turbo kicking in, and the vcore is increased to compensate.

My chip actually spikes to as high as 1.464V for a second or so, but most of the time it is in the realm of 1.424-1.456V during Turbo. At idle, the vcore is 1.184V.


My best undervolt settings (Turbo Off, CnQ off, stock 3.2GHz, all 6 cores enabled):
vcore @ BIOS: 1.157V
vcore (CPU-z, idle): 1.136V
vcore (CPU-z, peak@IBT/P95): 1.184V
Idle Temp: ~35-36C
Max Temp (IBT/P95): ~48C (using Thermaltake Frio OCK, 50% fan speed only)
Ambient Temp: within 32-34C range.

Perhaps this can help you in testing what your max undervolt is. From my experience, E0 stepping completely changed the voltage game from C3 (Deneb). I'm a little hazy, but my X4 965 could only go down to 1.24Vcore @ stock clocks.

You know what's even crazier? (And this is why my CnQ is turned off in my Max Undervolt setting). CnQ causes more voltage to be used at load, although it does have the benefit of cores being clocked down when idle.

For example:
Stock Settings, CnQ on, P95/IBT Max Vcore: 1.376V
Stock Settings, CnQ off, P95/IBT Max Vcore: 1.232V

It's crazy, so I just turned it off; anyway, when I'm running the thing undervolted, don't really care about Turbo (Turbo gets disabled automatically without CnQ).

 

[mikeymikec]

Please note that I said that under load the core voltage was lower than when idle I'm guessing it's because the readings aren't really in sync and if I want to monitor frequencies and voltages closely I ought to use something perhaps like ASUS's own monitor programs.

 

[KingFatty]

Also, changing DRAM settings may enable you to maximize the other parameters of the CPU (CPU undervolting, achieving a higher CPU clockspeed at a given volage, etc.)

Are you running some kind of server or memory-intensive business application? If not, I'd suggest turning down the DRAM settings to the lower setting, easing the burden on your CPU's integrated memory controller thereby allowing the CPU to run cooler thereby allowing you better flexibility with CPU clocks/voltages that will probably let you get away with even lower CPU voltage at the same clocks.

I just figure that the benefit you get from running the memory at 1600 is very small compared to running it at 1333, but the negative impact on the CPU is greater and will reduce your ability to undervolt at a given clock. Also, there is even a chance that your DRAM will require less voltage at 1333 further reducing power consumption and heat. I guess for ballpark guesstimate, lets say that using 1600 RAM gives you an increase in performance of 0.5% compared to 1333, but running the RAM at 1333 enables you to increase the CPU performance by 4% (either by enabling you to lower CPU voltage for a given speed, or by increasing the CPU clock speed for a given CPU voltage).

But I don't have your particular CPU, though I have the same motherboard, so YMMV. In my case, my RAM uses lower voltage at 1333, and I achieved a higher CPU overclock, when I switched from 1600 down to 1333 RAM speed.

 

[jvroig]

Oh I misinterpreted that a bit, sorry. I thought you didn't mean completely idle, only that when you haven't started prime yet (so some random background task is actually running, like after booting and programs that are loading trigger Turbo, making the vcore fluctuate a lot), I somehow connected it to the varying vcore you mentioned upfront and completely misread your second post mentioning the complete idle vcore.

 

[mikeymikec]

I've knocked the voltage down by 0.02v now, and the only odd thing I've noticed is that I don't think (although if I had been more scientific in my approach, I would have made notes of the figures throughout), the core voltage reading in CPU-Z has actually dropped by 0.02v. At the moment I'm seeing a voltage reading of 1.272 compared to 1.284 when I started, which is a variation of 0.012v (CPU completely idle when taking the reading).

Is it possible that my board is trying to be clever and over-ruling my setting on the fly instead of the system just locking up and requiring me to restart and reset the setting?

My earlier idea of trying to find a more accurate voltage reader worries me a little because on the ASUS forums they sometimes say "the stability issues cleared up after I removed ASUS whatever monitoring app".

Using SpeedFan's voltage reading as an additional indicator, it seems to agree with CPU-Z, though it doesn't show as many decimal places (2 rather than 3).

I've also set my RAM to DDR3-1333.

- edit - Huh?!? There it goes again, I'm running the 7z benchmark and core voltage has dropped to 1.23 - 1.24v while the processor is at virtually full load. Admittedly the voltage sometimes goes up to 1.4xv but a lot of the time the load voltage is lower than the idle voltage, according to CPU-Z and SpeedFan.

Admittedly I normally run Prime95 as a stress tester, but my first thought was the 7z benchmark.

 

[Hatisherrif]

- edit - Huh?!? There it goes again, I'm running the 7z benchmark and core voltage has dropped to 1.23 - 1.24v while the processor is at virtually full load. Admittedly the voltage sometimes goes up to 1.4xv but a lot of the time the load voltage is lower than the idle voltage, according to CPU-Z and SpeedFan.

Admittedly I normally run Prime95 as a stress tester, but my first thought was the 7z benchmark.

Happens to me with an i5 2500K too. As long as your system is stable, you will be fine.

 

[mikeymikec]

I've brought it down by 0.04v, and I'm just wondering, do people think there's any chance that the board may forget to re-apply the specified offset voltage and go back to standard settings after sleep/resume?

I'll restart now, but my computer seems to be idling (as in, lowest clock speed in use, CPU is idle) and the voltage is 1.296v.

- edit - restarted now, yes, the idle voltage shows up as 1.248v now. A BIOS update may be in order methinks.

- edit - BIOS updated. Immediately after restart, idle voltage shows up as 1.248v. After sleep/resume, and waiting for it to completely idle and for the voltage to return to normal, 1.296v.

- edit - as a test, I've disabled sleep mode for the time being and enabled hibernation. Just hibernated/resumed, idle voltage is 1.248v. I plan to contact ASUS in the hope that they can do a BIOS update fix for this issue (assuming that it is a BIOS-update-fixable issue).

 

[mikeymikec]

Hmm, my undervolting experiment has probably come to an abrupt end - the BIOS won't let me reduce the offset by anything further than 0.0625.

Note - it's not that the system won't start if I set a value lower than that, the BIOS simply won't allow a lower number in that field. It's stable on this setting seemingly (ran Prime95 for ten minutes and Orthos for 10 - the latter to invoke Turbo Core, both on small FFTs to stress the CPU).

Any ideas as to why it won't allow a lower setting? I would have thought that if it was a question of stability, then it would allow me to enter the setting, save, then reboot and complain that it doesn't like that setting, or reboot with that value reset to the default.

Latest BIOS version installed already.

 

[blckgrffn]

Any idea how much power at the wall you are saving?

Sorry, I don't have any help when it comes to tweaking the bios further...

 

[mikeymikec]

Not very scientifically done (as in, I got a quick reading in each state):

0.06v offset: 85-90W idle, 175W load
standard: 89-95W idle, 186W load

Personally I think the idle readings are probably going to be about the same on average. - edit - just checked the 0.06v offset idle reading again, it was 90W.

Load = Prime95, all cores, small FFTs
Idle = CPU graphs on systray are all flatlined, no sound of disk activity, core voltage where I would expect it to be when idle

Bit disappointed I hope someone has a suggestion that works.

- edit - I've checked the BIOS again, the information on the side for that setting actually says what the minimum and maximum offset values can be, and that's where it's stopping, so it's by design, the question is how can I get around it, if at all possible.

I've thought about temporarily disabling CnQ/Turbo Core, but I'm not sure that either would make any difference here. Also, I can't believe that if I disable CnQ that I could possibly get idle power usage down to anywhere near what is managed with CnQ enabled, so that seems like a bit of a non-starter. Doing a bit of googling atm.

 

[blckgrffn]

Not very scientifically done (as in, I got a quick reading in each state):

0.06v offset: 85-90W idle, 175W load
standard: 89-95W idle, 186W load

Personally I think the idle readings are probably going to be about the same on average. - edit - just checked the 0.06v offset idle reading again, it was 90W.

Load = Prime95, all cores, small FFTs
Idle = CPU graphs on systray are all flatlined, no sound of disk activity, core voltage where I would expect it to be when idle

Bit disappointed I hope someone has a suggestion that works.

- edit - I've checked the BIOS again, the information on the side for that setting actually says what the minimum and maximum offset values can be, and that's where it's stopping, so it's by design, the question is how can I get around it, if at all possible.

I've thought about temporarily disabling CnQ/Turbo Core, but I'm not sure that either would make any difference here. Also, I can't believe that if I disable CnQ that I could possibly get idle power usage down to anywhere near what is managed with CnQ enabled, so that seems like a bit of a non-starter. Doing a bit of googling atm.

Thank you - that's good enough for me

I would try disabling Turbo Core. You are out almost nothing and would be better served 99.9% of the time by just cheating and increasing the multiplier on the CPU up from 15 to 16 (I believe).