Ivy Bridge Voltage vs Temperature vs Power (With Charts!)

[Yuriman]

Reading another thread sparked my interest in undervolting, and since I hadn't played with it since changing motherboards and moving SFF, I decided to see how low I could take my 3570K.

Using a fixed clockspeed of 4ghz (a 400mhz/10% OC), I began lowering the offset and recording temperature as well as idle and load voltage and wattage (at the wall, with a Kill-A-Watt) using IntelBurnTest (AVX) to generate load. I did not use a fixed fanspeed profile. However, the fanspeed was ~2500rpm at the lowest offset, and ~2580rpm at the highest offset, so the impact of this should be negligible and it saved me time in collecting data.

My HD7850 is good for around 10w fixed power draw, and peripherals attached via USB total in at about 3.5w.

Using a +0.005v offset, I had a load temperature of ~92c on my highest core, ~129w total system power draw and a power delta of 82 watts at the wall.

Using a -0.130v offset, I had a load temperature of ~76c on my highest core, ~107w total system power draw and a power delta of 63 watts at the wall.

Going by Anandtech's review of my PSU (Antec Earthwatts 380w), the conversion efficiency should be around 85% at ~30% load, which is approximately where my 380w unit is during a CPU-only stress test. Efficiency should be at around 80% at idle/~10% load. This means that my total system power draw before AC-DC conversion losses is approximately 110w and 91w, with power deltas between idle and load of 77w and 55w before and after undervolting.

Taking peripheral and video card power draw out of the equation, the rest of my system draws around 93w under full load, and ~30.5w at idle.

Here's a link to my data.

 

[Yuriman]

For those curious, my cooler is a Noctua NH-L9i. Here's a photo stolen from google images:

 

[fuzzymath10]

Any idea how much lower you could go with voltage at stock?

I chose the wrong motherboard for my mitx i7 3770 (z77-n wifi) and I'm on the edge of maxing out the picopsu, even at stock because there is no way to reduce voltage. All I can do is set the max wattage and current in the bios and it hard to tell what changes in behaviour those introduce.

 

[Yuriman]

Unfortunately, I don't have an easy way to lower turbo-only voltage like I could have on my previous ASRock board. I only have a fixed offset, and eventually my idle and partial load voltages would get low enough that I'd experience instability.

How much lower? Not sure. 0.856v is pretty low for 1600mhz, by my estimation.

 

[BSim500]

How much lower? Not sure. 0.856v is pretty low for 1600mhz, by my estimation.

Good job Yuriman. :thumbsup: I have an i5-3570 and with a -0.10v offset it runs at 0.696v idle (1.6GHz) and 1.004v load (4.0GHz) with a delta of just under 55w. Temps never exceed 60c on a 212 EVO. Draws around 20w less than my old OC'd i3-530. I think that's towards the "winning end" of the "silicone lottery", and I've been highly reluctant to get rid of it.

Any idea how much lower you could go with voltage at stock?

I chose the wrong motherboard for my mitx i7 3770 (z77-n wifi) and I'm on the edge of maxing out the picopsu, even at stock because there is no way to reduce voltage. All I can do is set the max wattage and current in the bios and it hard to tell what changes in behaviour those introduce.

If you're having stability problems on your PSU, there's a trick that'll allow you to (almost) mimick an "i7-3770S" chip that should force a lower load power consumption:-

Control Panel -> Power Options -> Balanced -> Change Plan Settings -> Change Advanced Power Settings -> Processor Power Management -> Maximum Processor State from 100% to 90%.

This effectively tells Windows to use the second highest Speedstep setting as the maximum load state (testing it on mine it drops to 3.0GHz @ 0.832v at 31w delta) which is pretty much the same base clock as an i5-3330 but the low power of virtually an i3-"T" variant). No extra software is needed or need to reboot. Doing it this way disables your Turbo Boost but that's often not really an issue in a HTPC, and stability is far more important than turbo. It should also work on motherboards which lack undervolting control in the BIOS as it's part of the core Speedstep functionality rather than a manual undervolt.

 

[Yuriman]

This morning I did some testing with Windows' Maximum Processor State, leaving my -0.11v offset in place. Load voltage and wattage were measured the same as before, sitting on the desktop and running Intel Burn Test to generate an AVX load.

The voltage floor for my chip is ~0.880v on this motherboard as reported by CPUz, and would be ~1.000v without the offset, using the stock voltage table. Lowering maximum processor frequency gradually, it reaches this value at 2800MHz, or 84% maximum processor state.

CPU-only idle power draw is an unknown quantity, but if we assume approximately zero, the efficiency of the CPU continues to increase as frequency falls, despite voltage leveling out below 2800MHz. Realistically, I'd guess the CPU is responsible for ~5w of power consumption at idle.

However, if we factor in total system power consumption, it's most efficient to run the CPU just below its maximum non-turbo frequency (~2800-3100MHz).

In a system with more static power consumption than mine, peak efficiency would actually be right about max non-turbo.

Some food for thought: If we assume ~5w idle draw, under an AVX load and accounting for PSU efficiency @ ~80%, an Ivy Bridge desktop chip draws ~12w at 1600MHz, ~21w at 2600MHz, and ~36w at its maximum non-turbo frequency of 3400MHz. Going up another 600MHz increases power consumption by almost 50%, which makes it pretty clear why the stock frequency is what it is.


Here's a table with my data:
(note that P-state is actually Windows' Maximum Processor State, in percent)

 

[know of fence]

So trying to understand your last post, I have a lot of questions. But first why can't you just set your Vcore without messing with the offset. On my board lowest Vcore I can set is 0.800 V and only then can it be lowered further with the offset. Also at 800 MHz the G3258 goes all the way down to around 0.726 V, meaning that required voltage continues decreasing with every multiplier. You are nowhere near the threshold voltage for 22 nm, yet.
I figure that Speedstep is pretty generous with these low volages, at x15 I get 0.817 at x 17 i get 0.841. Curiously I can't seem to be able to set a multiplier of x16 using just the Windows Power Options (maximum CPU state %).
You managed to set a 4 GHz OC with a fixed voltage, can't you just set the same for 1.6 GHz and see how low you can push the voltage. Which might bolster your confidence setting bigger negative offsets while using all those wonderful power saving features.

 

[Yuriman]

So trying to understand your last post, I have a lot of questions. But first why can't you just set your Vcore without messing with the offset. On my board lowest Vcore I can set is 0.800 V and only then can it be lowered further with the offset. Also at 800 MHz the G3258 goes all the way down to around 0.726 V, meaning that required voltage continues decreasing with every multiplier. You are nowhere near the threshold voltage for 22 nm, yet.
I figure that Speedstep is pretty generous with these low volages, at x15 I get 0.817 at x 17 i get 0.841. Curiously I can't seem to be able to set a multiplier of x16 using just the Windows Power Options (maximum CPU state %).
You managed to set a 4 GHz OC with a fixed voltage, can't you just set the same for 1.6 GHz and see how low you can push the voltage. Which might bolster your confidence setting bigger negative offsets while using all those wonderful power saving features.

I seem to have missed this post! Better late than never to respond, though.


I understand that I'm nowhere near the threshold voltage for 22nm, but if I understand your question, using a fixed voltage rather than using the processor's voltage table will be strictly academic, as I don't want to run with a fixed frequency, either, and the most power savings will be with the least voltage at the highest frequency. I didn't use a fixed voltage for my 4ghz OC; rather, I found the lowest offset that was still stable at my 4ghz OC, and then scaled the processor's frequency down using Windows Power Options. It's likely I could run at even lower voltages at lower frequencies, but I can't effectively go any lower due to limited voltage control in bios.

 

[know of fence]

Overclocks aren't very power efficient, due to increasingly large Vcores they require. However I always believed that power efficiency would increase with lower clocks. With automatic Speedstep voltages this doesn't seem to be the case. A moderate 4.3@1.221 V overclock seems as power efficient as the low voltage states, which are also offset by [-0.169 V] and decrease to values as low as 0.550 V.

Unlike Sandy, Ivy Bridge and Skylake - Haswell/Broadwell CPUs have the ability to scale down voltages and frequency to very low states, but it seems to only benefit idle power.
Actually running OCCT on the CPU in those low states (800 - 2000 MHz) is anything but power efficient.

This graph pretty much mirrors your last one, it uses GHz for performance assuming that most common benchmarks would be proportional to clocks. Graph uses CPU reported "Package Power" from HWinfo64.

 

[Yuriman]

Thanks for your data. I'm skeptical as to the accuracy of package power numbers, as they don't seem to agree very well with what I derive from my wall current measurements, but at least they should be self-consistent.

A few days ago I collected a new set of data for my new power supply, and decided to remake my charts. Looks like moving from a 400w bronze-rated PSU to a platinum rated PSU ($25) shaved off around 10w under CPU-only load, and 5w at idle.

Total system power consumption at the wall, CPU AVX load @ 4ghz with decreasing offsets:

Delta watts (difference between CPU idle and load) at the wall at 4ghz AVX load as vcore drops:

Core temperatures as voltage/wattage drops, at 4ghz AVX load:

 

[Yuriman]

Using a -130mv offset, these are the voltages I get at various frequencies, effectively showing what Ivy Bridge's voltage table looks like:

Efficiency (frequency vs power consumption) of the CPU alone:

Efficiency (frequency vs power consumption) including total system power consumption:

It's worth noting that the ratio of CPU power consumption to system power consumption will shift this curve left or right - large "parasitic" platform power consumption will push things toward higher overall efficiency at higher frequencies, while platforms that sip power will show the greatest efficiency at lower clocks.

 

[Yuriman]

Here's a link to my raw data, and my spreadsheet that I used to generate the graphs, in case anyone else would like to use it:

https://drive.google.com/file/d/0B4C9i5e_Qs8IQ0RUbkM2a2lXYnM/view?usp=sharing

 

[know of fence]

Regarding package power accuracy, it seems to pretty much follow the curve at the same 12-13W difference. I use the max value for 2 min test (there is also average, and current power). Now this is a system with components selected for low power/noise. I was trying to see how close desktops can come to those dreadful mini-PCs, ~20 W idle is not so bad.

 

[Abwx]

T

Some food for thought: If we assume ~5w idle draw, under an AVX load and accounting for PSU efficiency @ ~80%, an Ivy Bridge desktop chip draws ~12w at 1600MHz, ~21w at 2600MHz, and ~36w at its maximum non-turbo frequency of 3400MHz. Going up another 600MHz increases power consumption by almost 50%, which makes it pretty clear why the stock frequency is what it is.

At stock and with Prime 95 it is 52W for the 3570K, so not sure what is your AVX load..

http://www.hardware.fr/articles/863-11/overclocking-undervolting.html

There s a comparison with previous gen, so far HFR provided the most accurate numbers before sites like THG took advantage of their bigger budgets to buy (or receive.?.) professional measurement gear.

 

[sm625]

Thanks for your data. I'm skeptical as to the accuracy of package power numbers, as they don't seem to agree very well with what I derive from my wall current measurements, but at least they should be self-consistent.

A few days ago I collected a new set of data for my new power supply, and decided to remake my charts. Looks like moving from a 400w bronze-rated PSU to a platinum rated PSU ($25) shaved off around 10w under CPU-only load, and 5w at idle.

Total system power consumption at the wall, CPU AVX load @ 4ghz with decreasing offsets:

Delta watts (difference between CPU idle and load) at the wall at 4ghz AVX load as vcore drops:

Core temperatures as voltage/wattage drops, at 4ghz AVX load:

These images dont load.

 

[Yuriman]

Sorry, will relink.

EDIT Fixed.

 

[Yuriman]

At stock and with Prime 95 it is 52W for the 3570K, so not sure what is your AVX load..

http://www.hardware.fr/articles/863-11/overclocking-undervolting.html

There s a comparison with previous gen, so far HFR provided the most accurate numbers before sites like THG took advantage of their bigger budgets to buy (or receive.?.) professional measurement gear.

What exactly is the question? I appreciate the link to more data.

I use the latest version of Prime. IBT's power figures weren't significantly different.

At stock frequency and undervolted (1.016v), my whole system draws 81.6w at the wall with Prime95 running, which gives a delta wattage (difference between idle and load) of 44.4w. If the CPU alone is drawing 5-10w at idle, this is in-line with a ~52w power consumption figure - though I've measured mine at the wall.

I'm pretty sure my overall parasitic power draw is a lot lower than most 1155 systems'. I dropped almost 40w at idle when I moved from my ASRock Z77 full-ATX to my Zotac Z77 ITX board.

^ 42w on the 12v rail is reasonably near to my own numbers.

 

[Yuriman]

Update, now with overclocking.

I scaled my i5 from 1.6ghz to 4.5ghz, using the minimum stable voltage at each step, and found the minimum wattage/frequency curve of this 22nm CPU. Note that I was not able to lower my offset below -150mv because I got crashes at idle/partial load. If I was able to get into Windows and get Prime running at, say, -180mv, it wouldn't crash, but if I walked away without a stress test going I'd come back to a frozen cursor. 3.8ghz was very happy/stable at -150mv, so that's probably where I'm going to leave my system at now.

And, a continuation of the efficiency curves:

Efficiency of the CPU alone: