Power Consumption Scaling with Clockspeed and Vcc

[Elephantjoli]

[/URL][/IMG]Hi everybody,

I've been looking for a way to estimate my CPU power consumption and, of course, I've found that on the forum.
Here is the link to this great post by Idontcare : https://forums.anandtech.com/thread...-clockspeed-and-vcc-for-the-i7-2600k.2195927/

I've more or less follow what he has done and apply it on my computer.
First I held the operating voltage while increasing the clock speed. I've done it for lots of voltage.
Here are my results :

As you can see, the consumption is linear with the clock speed (I've avoid the clock speed higher than 3300 MHz for further analysis). I haven't represented the equations on the graph so it can be clearer.

Still following Idontcare's work, I've plotted the interception with 0 GHz for each voltage.
Here is the graph I got :

The result is supposed to be linear and .... it's clearly not here. The tendency equation on the graph is 2,2*Vc^6.75+65.

Could someone help me with that ? What's wrong with me ?

Btw I'm working on a AMD FX 4300 and my mb is a Crosshair V Formula-Z.
I'm also sorry for my shitty english level, I'm French. Please pardon my French ...

 

[Abwx]

What matters is the power delta between idle and load, total power consumption alone is irrelevant.

 

[Elephantjoli]

What matters is the power delta between idle and load, total power consumption alone is irrelevant.

Yeah it's true if you want to have an idea of your CPU performance. I'm sorry if I haven't been clear enough but this is not what I'm looking for.
I'd like to find out what part of my total power consumption goes trough my CPU and what part doesn't. I'd like to get something like that for my own computer :

This would be very useful for the thermodynamique application I'm working on. I definitely need to know how much power goes in the CPU and how much in the other parts of my computer.

 

[Andrei.]

The result is supposed to be linear and .... it's clearly not here.

The result is exactly what it's supposed to be. Power consumption if P = V² * f * C + V * t * k

Ignoring static leakage, power scales V² * f. That's why you see a linear curve with varying frequency and a quadratic curve with varying voltage.

 

[Elephantjoli]

The result is exactly what it's supposed to be. Power consumption if P = V² * f * C + V * t * k

Yes to be exact the power consumption of a CPU is :

But when you add the consumption of the CPU and the other components you get the following equation :

So when every interception for 0 GHz are taken, a linear result is supposed to be found :
P_total (0 GHz) = A + m*Vcc

After that the interception for Vcc=0 is supposed to give the consumption of the computer outside of the CPU.
For sure the post by Idontcare is clearer than I am : https://forums.anandtech.com/thread...-clockspeed-and-vcc-for-the-i7-2600k.2195927/

 

[Andrei.]

Yes to be exact the power consumption of a CPU is :

But when you add the consumption of the CPU and the other components you get the following equation :

So when every interception for 0 GHz are taken, a linear result is supposed to be found :
P_total (0 GHz) = A + m*Vcc

After that the interception for Vcc=0 is supposed to give the consumption of the computer outside of the CPU.
For sure the post by Idontcare is clearer than I am : https://forums.anandtech.com/thread...-clockspeed-and-vcc-for-the-i7-2600k.2195927/

I don't understand where those values are coming from. To measure static leakage you are supposed to vary voltage at idle while disabling power gating states, you seem to be varying voltage at load as that's exactly the result of your graph.

Also that formula is wrong. Idontcare is totally ignoring temperature and it has a massive impact on high performance nodes such as used on desktop CPUs. The whole Vcc^B thing is nonsense, it's Vcc² always. Use P = V² * f * C + V * t * k and ignore everything else.

Edit: His formulas in this post are more accurate: https://forums.anandtech.com/thread...e-ghz-and-power-consumption-analysis.2281195/ but still don't agree with the methodology to measure static.

 

[Borealis7]

You dare to question IDC?! blasphemer!
but yeah, total power consumption is mostly impacted by V^2, by Frequency and Temperature. the rest is pretty constant in a single system.

 

[Elephantjoli]

First thanks for your answer Andrei.

Also that formula is wrong. Idontcare is totally ignoring temperature and it has a massive impact on high performance nodes such as used on desktop CPUs. The whole Vcc^B thing is nonsense, it's Vcc² always. Use P = V² * f * C + V * t * k and ignore everything else.
.

Your totally right on this point, temperature has a massive impact on consumption. That's why I've done all my test at a constant temperature (more or less 2°C). I guess IDC was aware of that too, just see the post he made on that if you doubt it : https://forums.anandtech.com/thread...-power-consumption-with-the-i7-2600k.2200205/

I don't understand where those values are coming from. To measure static leakage you are supposed to vary voltage at idle while disabling power gating states, you seem to be varying voltage at load as that's exactly the result of your graph.

If I do vary voltage at idle (with a constant clock speed ?), what would I be supposed to get ? I'm asking cause IDC give different formula for the static leakage, one fonction of V and one of V²*T. I must admit that it's puzzling me.

Still don't agree with the methodology to measure static.

What would be your way of doing it ? I mean being able to say from the total power consumption : this part goes in my CPU and this part goes in the other parts of my computer

 

[Andrei.]

If I do vary voltage at idle (with a constant clock speed ?), what would I be supposed to get ? I'm asking cause IDC give different formula for the static leakage, one fonction of V and one of V²*T. I must admit that it's puzzling me.

Static leakage is V*t*k where k is some coefficient depending on process and design. Make sure temperature is probed on the core sensor. IDC's latter formulas are alright, it's just the first one with Vcc^B that you first quoted which is nonsense.

What would be your way of doing it ? I mean being able to say from the total power consumption : this part goes in my CPU and this part goes in the other parts of my computer

Did you verify that the CPU is in a clock gating state and power gating is disabled? It looks like it's not clock gated. Either that or the voltage regulators are utter shit.

 

[Elephantjoli]

Static leakage is V*t*k where k is some coefficient depending on process and design. Make sure temperature is probed on the core sensor.

When I do that, I get something linear (more or less) with the voltage but which intercept V=0 with a negative power value (-26 W). How can this be possible ?

Did you verify that the CPU is in a clock gating state and power gating is disabled? It looks like it's not clock gated. Either that or the voltage regulators are utter shit.

I'm not sure to understand this. Do you mean that I must disable all the option in the bios such as "cool'n'quiet", "power saving mode", etc ?
Btw, I use a voltmeter to check the voltage regulation, no problem on this side.