Indirect CPU power measurements.

Mr Evil · Post by **Mr Evil** » Tue Jun 19, 2007 11:29 am

CPU power consumption is something that's talked about a lot here, but it's not easy to measure directly. The approximate (very approximate!) power consumption of the CPU can instead be calculated indirectly from the total system power consumption which can be measured more easily with a mains power meter (yeah, I've just been looking for excuses to use the one I bought recently). CPU power consumption is approximately proportional to the square of the voltage, so by measuring the power consumption at various voltages, the difference can be discovered.

The test system can be seen on a page on my website. I performed the measurements running Prime95 with the small FFT size so that components other than the CPU are affected as little as possible. I used RMClock to vary the voltage. All tests were run at the default frequency of 2.1GHz. The raw numbers were:

Code: Select all

V/V        P/W
1.1        98
1.2        102
1.3        108
1.375      113

Going over 1.375V with RMClock involves registry editing, so I didn't bother.

Plotting P against V^2 gives a fairly straight line, as hoped:

I've shown the graph extrapolated down all the way to 0 voltage. There's no telling if the graph would actually continue straight like that, it's just there to show the intercept of the best-fit line.

The equation for the best-fit line has handily been calculated for me (there was a time when doing this would have involved actual maths - now it's all automatic). The number of interest here is 71 which is approximately the contribution to power consumption of everything apart from the CPU. The actual CPU power consumption is then 27W at 1.1V, rising to 42W at 1.375V. These numbers are in line with what I would expect given where my A64 X2 4000+ lies within the range of CPUs given a 65W TDP by AMD.

Note that the power consumption figures all include the inefficiency of the ATX PSU and the CPU's voltage regulator. If both of those were 85% efficient, then actual power consumption would be 0.7x lower.

EDIT: I forgot to say that I only ran one copy of Prime95, so only the FPU of one core was under load. I suppose I should have loaded both cores.

jaganath · Post by **jaganath** » Tue Jun 19, 2007 1:03 pm

if you have the time it would be great if you could run the experiment again with 2xPrime95. Also it's very unlikely either the PSU or VRM are 85% efficient at that power draw, so the CPU likely draws ~20W at 1.1V, which is excellent; at 0.975V, which people have reported w/ Brisbane cores, passive cooling should be a piece of cake.

Mr Evil · Post by **Mr Evil** » Tue Jun 19, 2007 2:08 pm

Redone with both cores loaded:

The gradient is almost but not quite twice as steep as for one core (40 vs 22). This makes the power [email protected] and [email protected]

hmmm... is it possible to tell the forum to not automatically parse those numbers as email addresses? I can't seem to see a button for it.

jaganath · Post by **jaganath** » Tue Jun 19, 2007 3:35 pm

Looks like the TDP is pretty bang on in this case. I should have known 20W for a dual-core @ 1.1V was too good to be true.

is it possible to tell the forum to not automatically parse those numbers as email addresses? I can't seem to see a button for it.

put a space either before or after the "@" sign, that should do it, e.g 42W @1.1V.

prodeous · Post by **prodeous** » Thu Jun 21, 2007 12:01 am

Intresting results.

But looking at the available data its hard to judge if its a strait line or a curved line.

Is it possible to run more samples?

But i have to say, this approach is quite intresting in checking power consumption of the CPU. Well done.

Mr Evil · Post by **Mr Evil** » Sat Jun 23, 2007 12:16 pm

prodeous wrote:...But looking at the available data its hard to judge if its a strait line or a curved line...

It does look suspiciously like it flattens out at low voltage, possibly due to changing voltage regulator efficiency. If I have some spare time I'll try to take more measurements. I can test higher voltages, but unfortunately not lower voltages since that's the lowest it is stable at.

mcoleg · Post by **mcoleg** » Sat Jun 23, 2007 2:41 pm

Mr Evil, nice experiment.

btw, might you get to more vcore settings from bios? could be a bit easier that way.

is power saving setting disabled? c'n'q and such?

Felger Carbon · Post by **Felger Carbon** » Sat Jun 23, 2007 4:59 pm

jaganath wrote:...it's very unlikely either the PSU or VRM are 85% efficient at that power draw.

This is true. But it's the incremental efficiency that's at question.

If there's no load whatever on the PSU, the capacitive losses from switching are the same as they are at 100% load. That doesn't change. Much (all?) of the magnetic core losses are also the same at no load as at 100%. Consider: at no load at all, the PSU is 0.0% efficient! And yet, at 20% of maximum load, the PSU efficiency is getting well above 0.0%!

Obviously, the incremental efficiency is then much greater than 0%, and much greater than the total efficiency at 20% of maximum load

cmthomson · Post by **cmthomson** » Sat Jun 23, 2007 5:28 pm

Most power supplies (and VRMs) achieve maximum efficiency at about 75% of rated load. They are slightly less efficient at full load, and much less efficient under light load.

Most PC-class motherboards have rather inefficient CPU VRMs (which is why they run hot and need heat sinks); 80% is very generous indeed. Some workstation-class motherboards (eg, P5W, DQ6) have 8 or 12 phase VRMs that are much more efficient, but still well below 90%.

Nonetheless, this is interesting empirical data, if only to confirm that the vendors are being relatively honest...

mcoleg · Post by **mcoleg** » Sat Jun 23, 2007 6:13 pm

btw, Mr Evil, do you count in the vcore droop?

Mr Evil · Post by **Mr Evil** » Sat Jun 23, 2007 7:14 pm

mcoleg wrote:Mr Evil, nice experiment.

btw, might you get to more vcore settings from bios? could be a bit easier that way.

is power saving setting disabled? c'n'q and such?

Higher voltages are available from the BIOS, yes, but it would take ages having to reboot between tests. Using RMClock I could change the voltage very quickly. RMClock should also allow higher voltages, I just have to get around to applying the necessary registry setting.

Power saving was disabled.

cmthomson wrote:...Most PC-class motherboards have rather inefficient CPU VRMs (which is why they run hot and need heat sinks); 80% is very generous indeed...

I thought I remembered hearing that they were inefficient. Disappointing in a way, but on the other hand it means that real CPU power consumption is likely to be significantly lower than the values above. I also checked a review of my CoolerMaster iGreen PSU for its efficiency, and it's a bit less than 80% at the loads in the tests, which also lowers the estimated real CPU power.

mcoleg wrote:btw, Mr Evil, do you count in the vcore droop?

Not as such. I checked the reported voltage under load using SpeedFan and it was close enough to the set voltage that I used the set voltage for the numbers. It doesn't really matter if the actual voltage was more or less than that, as long as the relative change in voltage between tests was the same, and I'm not about to go sticking probes on my brand new motherboard and risk shorting something out (I wouldn't trust the motherboard sensors for more than a rough value)!

mcoleg · Post by **mcoleg** » Sat Jun 23, 2007 7:31 pm

i see...

still, that board looks like it will droop a fair bit under load; m/b sensors are not too accurate but they will show the difference, at least. the droop might account for that bit of a curvature (a possible curvature, that is...)