CPU Power Draw, Voltage or Frequency?

[Ch0z3n]

As far as I can tell both frequency and voltage effect the total power draw of a processor. Is one more important than the other? Like, would there be a significant different between like a 2ghz @ 1v and 1ghz @ 1v? Or would there big a bigger difference between 2ghz @ 1v and 2ghz @ 1.3v? Basically, is there some, hopefully not too convoluted, formula to calculating power draw? Or at least approximately.

[tehcrazybob]

Voltage has a much more significant effect than frequency. There's no simple equation to calculate direct power draw, but there is a simple equation to calculate changes in power draw:



As you can see, power changes linearly with frequency but quadratically with voltage. Please bear in mind, when you're using this equation, you don't want to start with Thermal Design Power for most chips; it's better to find a review with an actual power measurement. If this isn't possible, use the TDP but fill in the stock values not for your chip, but for the fastest chip in the range - it'll be more accurate.

[Ch0z3n]

How do I find out how much my chip uses? I know it is usually lower than the TPD, but not sure by how much. I have a Kill-A-Watt if that will somehow be helpful.

[QuietOC]

As far as I can tell both frequency and voltage effect the total power draw of a processor. Is one more important than the other? Like, would there be a significant different between like a 2ghz @ 1v and 1ghz @ 1v? Or would there big a bigger difference between 2ghz @ 1v and 2ghz @ 1.3v? Basically, is there some, hopefully not too convoluted, formula to calculating power draw? Or at least approximately.

Power is linearly proportional to frequency (at a fixed voltage)
Power is proportional to voltage squared (at a fixed clockspeed)

The funny thing is that the voltage to reach stability is also linear with clockspeed. So power tends to be proportional to clockspeed cubed at minimum stable voltages.

Voltage tends to matter much more than clockspeed.

My Sempron LE-1250 + 740G
800MHz @ 0.784V -- 18W Idle / 19W load
1.6GHz @ 0.784V -- 19W Idle / 22W load

[RedAE102]

How do I find out how much my chip uses? I know it is usually lower than the TPD, but not sure by how much. I have a Kill-A-Watt if that will somehow be helpful.

If you're referring to the Brisbane Athlon X2 5400+ BE in your signature, at stock voltage, your chip probably draws very close to the rated 65W TDP at stock VID. There is only the 5600+ at 2.9GHz above it in the 65W Brisbane line; the 3.0 and 3.1 GHz models are rated at 89W TDP. I've done some in-depth analysis and calculations, and my non-BE 5400+ at its stock VID of 1.325V (CPU-Z: 1.296V-1.312V) runs almost exactly its family-rated 65W TDP. It undervolts to 1.225V (1.200-1.216V CPU-Z), which gives me ~55W TDP, both by the formula and by my own measurements using the efficiency curve of my PSU (many thanks to SPCR for the numbers on that!).

Come to think of it, I did do some limited measurements with my old G1 3600+ 1.9 GHz to see if my findings validated the the formula, and indeed they did. I posted the results somewhere in this forum.

[Ch0z3n]

My 5400+ BE runs at 1.35v stock, my highest setting on CPUID is 3ghz 1.325v. So like a 4850e running at 1ghz @ .8v would use an absurdly small amount of power?

[RedAE102]

My 5400+ BE runs at 1.35v stock, my highest setting on CPUID is 3ghz 1.325v. So like a 4850e running at 1ghz @ .8v would use an absurdly small amount of power?

Actually, even your chip would have an 8.15W TDP at that speed and voltage. My non-BE 5400+ runs 1.2GHz @ 0.775V, 2.0 GHz @ 0.900V, and 2.8 GHz @ 1.225V. I could go lower at 1GHz theoretically, but only some of the 45W Brisbane chips, as well as the Tyler Turion X2/Athlon X2, which is Brisbane's mobile counterpart, can actually select voltages lower than 0.775V, so for idle, I just run the highest speed I can with the lowest possible VID. I would think yours could also handle 0.775V @ 1.2 GHz, or at least at 1GHz.

[Ch0z3n]

.8v is the lowest CPUID will let me choose, maybe its the lowest my mobo supports?

[RedAE102]

.8v is the lowest CPUID will let me choose, maybe its the lowest my mobo supports?

It requires a little hack... in your CrystalCPUID folder, open CrystalCPUID.ini in notepad, and change this line:
CqMinVID=30
to
CqMinVID=31

Close the file, saving changes, and close and reopen CrystalCPUID, and voila, you're able to use 0.775V! Just to make sure, you can verify it in CPU-Z.

[tehcrazybob]

If you want to determine roughly how much a chip consumes at stock settings, you can calculate that using the equation I posted and the numbers for the fastest chip in the same series - just put the top-end chip in for the Stock values, and your frequency in for the Actual values. This is because any random chip can technically be viewed as an underclocked version of the fastest chip.

[Aris]

i was always under the impression the clock speed of your processor doesnt effect power drain. That the only reason overclocking draws more power is because you have to up your voltage to maintain stability, and the reason underclocking draws less power is so you can drop your voltage.

in short, ive always been told Voltage is what determines power draw, not the clock rate of the processor.

[tehcrazybob]

i was always under the impression the clock speed of your processor doesnt effect power drain...

If you look at the equation further up, and try a few sample cases, you'll find that that's almost true. Raising the clock speed does bring the power demands up slightly, but it's a small difference. Clock speed changes don't alter power consumption nearly as much as voltage changes.

[plympton]

I ran some tests today to get my head around speed/load/power draw. On my HTPC, I have CrystalCPUID running 3 frequencies (1, 2, and 3 GHz) with 3 voltages (.8, 1.1, and 1.275 respectively).



I ran each of the frequencies at idle and load, and was pretty surprised by the results - load vs. idle had a MUCH greater impact on power usage than does voltage. At equal voltage (1.275), there is a 2 watt difference between frequencies (86, 88, and 90 watts), so frequency has a really small impact on power draw.



It was pretty amazing to see that a full-loaded Prime95 run consumes so much power vs. unloaded. You could say, "Duh!", but why? It must point to the fact that modern processors are REALLY good at shutting down their internals when not needed. I can see my next build perhaps running stock clocks and voltages the way things are going these days!

So, are things like undervolting and underclocking going to be a thing of the past??

Now that I think of it, it would be almost as good to idle at the lowest stable voltage, and just ramp up the voltage as load increases, no? It's probably tough to figure out what 3 GHz idle stable is, and nigh impossible to guess what "medium load" stable is, eh? Though if I could idle at 3 GHz at 79 watts, wouldn't it eliminate some of the "jerky performance" that some people say they experience when Cool'n'Quiet is enabled?

PS: The point's moot: I can't even idle at 1.200v @ 3 GHz, much less 0.800v!

-Dan

[QuietOC]

It must point to the fact that modern processors are REALLY good at shutting down their internals when not needed. I can see my next build perhaps running stock clocks and voltages the way things are going these days!

Supposedly the Core i7 can acutally completely shut current off to most of its cores. They aren't idle; they are flat off.

I've noticed the Wolfdales respond completely differently to the Brisbanes to voltage and clockspeed. I am sure the proportionality still works the same (xV^2, yF), but the factors (x and y) are much different. My E5200 system idles around 50W almost regardless of clockspeed or voltage. I might as well run a big voltage and overclock then. Edit: I just set it to 12x320MHz = 3840MHz @ 1.520V 129W load, 1920MHz @ 1.536V 57W Idle.

Wolfdales seem really inept at operating at low wattage--maybe it is the whole huge cache, power hungry FSB arragement? The <10W ULV mobile versions all have reduced size cache and reduced FSB. My Atom N270 (same process tech and FSB design) equipped Eee PC idles at same Wattage at 6x166MHz as at load at 12x83MHz.

[plympton]

Supposedly the Core i7 can acutally completely shut current off to most of its cores. They aren't idle; they are flat off.

That's interesting- isn't that something that AMD has promised for a while (independent C'n'Q for each core), but I think still has to deliver. Maybe on the newest Phenoms?

I wonder how modern OS's will mess this up, though? Something about task schedulers shunting processes off to low-use processors would surely have to take this into account, no? I mean, if you have a medium load core, 3 off, and a new background thread - do you choose to send it to a shut-down processor for speed, or stick to the single proc for less power?

All this makes me happy I bailed on chip design after a year - makes my head hurt.

-Dan

[Tobias]

What I don't understand is why that formula up there isn't ever used to calculate the power usage of the processor. Maybe I'm just stupid and don't understand what I'm missing, but to me it seems to be possible to get an approximate fix. After all, we have 2 unknown (PA and PS) and two formulas to work with...

PA/PS = FrequencyA/FrequencyS*(VoltageA/VoltageS)^2

And running two full load tests at different frequency/voltages will yield a Power Draw difference. In my case (I'm running an old 90nm Windsor 4200 rated at 89W TDP), running @stock I have a max powerdraw of 135W (2 instances of Prime95) and at [email protected] I get 71W. So PS = PA + 64.

From the above formula we get that PA = 0.206PS and substituting for PA yields 0.206PS + 64 = PS or 64 = 0.794PS, PS ~80.6W

Sure, due to measurement errors it is not very exact, but still a rather good clue no? And if someone would happen to have a Kill-a-Watt which measures in tenths of a Watt, then exactness would increase.

[yuu]

hello

you have total power 135, at 85% psu, it is 114 watt
idling at 71 watt with 80% efficiency ~ 55 watt

40 watt out of 50 is motherboard + hdd + ddr2,

you should be comparing 74 and 15 watts, the cpu alone,

74 watt 2200 Mhz / 1.3V
15 watt 1200 Mhz / 0.8V

2200/1200 . 1.69/0.64
1.83 . 2.64 ~ 5

15*5, that is 75 watts

at stock there is also the thing of vrm's and cpu heating much more and being slightly less effective.

[Tobias]

It is actually worse than that. I am currently running an old modded Zalman 400B (with terrible values on the rails) which was meassured at 70-72% efficiency at these loads by MikeC in his review a few years back. It was actually rather good numbers back in the days, but those days are long gone. Anyways, my point is that the difference in efficiency is small enough to be negligible. The final value will be much more influenced by the lack of a decimal on the kill-a-watt anyways...

If one is going to do this calculation one needs to use numbers from the same "state" or one will over shoot. 71W is for load, my idle number was 65, but you are correct that I didn't include efficiency of the PSU in my calculations. Power efficiency of the mobo etc one have to live with. It's impossible to calculate anyways:) Adjusting for efficiency the CPU power draw it would mean 56W power draw of the CPU + power circuitry on the mobo. Concidering that 89W TDP is for the processor family, that is completely in line and 2*Prime95 is not the same as TDP...

The point I was trying to make was that it is possible to calculate how much of the power drawn at the wall that is due to the CPU, rather than to guestimate the power draw of mobo and gfx and hdd etc. The only weakness of that method is measurement errors, although the effect of those decrease as the power draw increase. And if the efficiency of the PSU is known it is also possible to calculate the actual powerdraw of the CPU...

Btw, yesterday I switched mobo and grafics card for a mobo with IGP and shaved off 20W at idle (Now I'm down to 45@idle, it would be really interesting to know the efficiency of the PSU at these levels:)

[austinbike]

That's interesting- isn't that something that AMD has promised for a while (independent C'n'Q for each core), but I think still has to deliver. Maybe on the newest Phenoms?

AMD has this feature (independant dynamic cores) in the Opteron line, starting with Barcelona (65nm).

I would assume it is in Phenom as well because they leverage so much of the same core elements.

[austinbike]

How do I find out how much my chip uses? I know it is usually lower than the TPD, but not sure by how much. I have a Kill-A-Watt if that will somehow be helpful.

Some may disagree but I think that this is a crazy exercise. In my opinion people should be fixated on the total platform power draw (at wall power) and not trying to nail one component.

[yuu]

there is a lot of vrm, cpu efficiency loss, 1.55v and 0.85 : 1600mhz. its a mistery to me, its strange.

guestimated power draw of mobo and gfx and hdd etc is 55W
8800GT
P31-DS3L
40GB
2*2GB

CPu E2140 1600/3200




according your formula Tobias

[austinbike]

Here is how you figure all of this out:

[yuu]

using "Energy check 3000" at the wall the numbers are:

the vrm efficiency is so weird, because it is cpu efficiency, given that 3200Mhz overclocked raises the fsb bandwidth consumption related chipset and memory too. and that results visibly even better efficiency, how about that.




the 70watt (3200 1.35) and 15 watt (1600 0.85) numbers have been verified on another Ep45 board using the Energy saver windows tool