How to measure CPU power?

[mviara]

Hello. This seems a basic question. But I have not been able to find specific answers on Google and even on Tom's Hardware. Can anyone point me to an article that describes how the CPU (not the entire system) power can be measured?

I understand that this could be difficult to do, as one may need to insert a sense resistor before the CPU. Tom's Hardware must have a good experimental setup as they regularly report power numbers for various CPUs. I have not been able to find the article in which they report their measurement methodology - can anyone point me to it?

I have also been looking for boards that come equipped with such sense resistors. One such board is the Radisys Endura LS855 that uses a Pentium M processor. However, I need to use a dual-core processor for my experiments, which deal with dynamic thermal management for multi-core CPUs, so the Radisys board would not be suitable. I found this Intel article that uses a board called Jamison Canyon CRB that has a Core Duo is already instrumented with sense resistors for measuring CPU power. It seems like it is an Intel internal board, does anyone know more about this, or similarly instrumented boards?

 

[sithscout80]

Knowing the Vcore doesn't really help at all because the current is what changes (that is why more power is consumed when running the processor at 100% than when at 0% even though the processor still runs at the same Vcore)

You will need the Vcore measurement (along with the current) if you are wanting to measure the power management features on new cpus.

 

[vikris]

Measuring current that cpu draws at a constant or nearly constant voltage is very difficult, because you need a special equipment. You must go to a well equiped lab for that kind of experiment. Also you must be very experienced.
Maybe I'm wrong, but I doubt that this can be done in a home environment.

 

[mviara]

vikris, this is for an experiment to be conducted in a university lab. We are prepared to purchase the required equipment. Unfortunately, we are not experienced in this matter. I was hoping to find a link to how Tom's Hardware does this.

 

[13thmonkey]

I think that the way that THG does it is by comparison, i.e. compared systems at idle to systems at full. if you do this with certain specialised applications, prime with little memory for instance, you can isloate the CPU to a degree but that only tells you the rise to a higher level. But most tests appear to be consistent system at idle compared to full, i.e. to test a GPU slap a full load on GPU with consistent CPU loads across all test platforms, and measure the changes in wattage to see which is more efficient a low or high loads.

It might be possible to surmise the idle wattage if you know the max electrical power of a CPU (this would be higher than the TDP else it is a perfect heating machine, the perfect CPU would use a lot of electrical power but create very little heat in comparison.)

Say for instance that the max electrical power was 100W. And we know that going from idle to full power (CPU intensive only) creates a delta of 60W in our test rig. We could estimate that at idle the CPU was drawing 40W. These figures have no basis in reality as I don't have such figures but the theory is about right. The only trouble is how do you fully load a CPU, and could they actually cope with being fully loaded?

Just a thought as to how you could get an estimate.

 

[exit2dos]

Not sure about THG, but here is how X-bit does it:

Our method remains unchanged. We measure the consumed current (that goes along the 12V CPU power circuit) with a clip-on ammeter. The real voltage in the circuit is verified with an ordinary multimeter. Based on the fact that the CPU is fed by the PSU through a separate circuit, this method is not very accurate since the CPU power converter is not 100% efficient, yet its accuracy is satisfactory for our purpose, especially since the power consumption of processors may vary a little from sample to sample.

http://www.xbitlabs.com/articles/cpu/display/sempron-3000_9.html

 

[croc]

Good question...

Since you have a lab at your disposal, I would think that its relatively easy to answer, though... Me, I'd probably start out with a good, calibrated clamp-on ampmeter around the AC to measure the total draw. Then I'd probably find the various +- leads for +5VDC, -5VDC, +3.3vdc, etc. and get some high current rated variable resistors, also preferably calibrated for resistance. I'd put some (calibrated) high load capable ampmeters in line to measure the effective load for each voltage. Then take the amperage of all leads at infinite restinance, vs the amperage of the AC, that will be your base load. Or effeciency at idle. now start dialing down your VR's, either one at a time or all together... Don't know what it is your trying to achieve, so you'll have to design your experiment yourself. But the setup described should get you close to where you want to go.

Be sure you have a LOT of psu's.... ;-)

Good Luck!

Edit... Sorry, thought you were trying to measure PSU's... OK, just one ampmeter on the +3.3VDC output should do it...

 

[mviara]

Though this appears simple, I feel it may be a very good first cut. However, it may not be accurate enough to report the results in an academic journal, so I will have to eventually get to the CPU directly.

 

[13thmonkey]

you want academic jounral quality?

Tricky...

Given that all of the power comes from the 3phase/4phase/8phase smoothing caps, would it not be possible to 'cut' the legs of the caps, mount them remotely so that you can find the current that they are passing, the only problem with this is that you'd probably affect the current that they are feeding the CPU and perhaps change it performance characteristics.

Else do what X-bit does and use a clamp-on ammeter? not sure of the accuracy though.

 

[audiovoodoo]

I like the simplicity and low cost of that approach but unless you connected to the circuit track would your results not be badly affected by the efficency of the power stage on the motherboard? you could see difference between two processors on the same motherboard but you would be way off what the CPU alone was using.

Have you tried calling Intel and AMD and asking them? I would have throught they had lots of ties to accademic institutions and this sort of information is hardly going to fall under NDA.

edit: Think 13thmonkey beat me to it there..

 

[mviara]

exit2dos, thank you for your reply.

I must admit I haven't looked at the inside of a PC in a while, especially after switching to the Mac. I read about the different supply rails in a PSU here. From what I understand, recent dual-core CPUs have an 8-pin plug from the CPU that exclusively powers the CPU. Measuring the current through this rail would then give a quite accurate estimate of the CPU power.

I would like to measure the power consumption at small intervals of time, so I guess I'll have to bring in a sense-resistor and DAQ eventually, but the clip-on ammeter would be a good start.

 

[13thmonkey]

the 6/8pin aux socket does solely supply the cpu, but its not the only supply to the cpu.

So checking that connection may only give you some of the amps

 

[mviara]

Hmm. Is that why Xbitlabs says their method may not be accurate? This makes the power measurement really difficult. But how come Xbitlabs, Toms Hardware, etc. present their power comparisons with such confidence? Perhaps because they are only interested in relative power numbers. Even then, as PSUs differ among CPUs, the error would not be the same across all of them, so it would even affect relative numbers, wouldn't it?

 

[13thmonkey]

if you take power into the PSU (v Simple) and keep as much as possible constant you can get a power reading at various levels of activity.

If you swap AMD for Intel you can get power readings for the platform at idle, if you only swap the mobo and chip. you have a platform to platform comparison. Which is a real-life comparison.

Any other figures are either clamp-ons or are based on electrical power specifications (not Thermal power)

 

[mviara]

AudioVoodoo, as I mentioned in my original post, the Jamison Canyon CRB board from Intel mentioned in an Intel article would be all that I need. I am trying to contact them for this. I actually did an internship with them in the summer, but could not do much as I did not get permissions for various tools. So, it may not be easy to get this board. Also, this particular board is not publicly advertised as an evaluation platform, so it becomes a matter of having the right contacts in the right group at Intel.

 

[croc]

That's why you go through the whole bit with the PSU first... you need to establish a baseline, on all the circuits, of the PSU. Then knowing those numbers you can compare the 3.3 going to the CPU in isolation. Might help to leave the other ammeters in place to see what effect the CPU drawing power would have on other legs...

Only difficult bit is something that purely stress the CPU, and in various loads. SuperPi would stress the CPU quit a bit, but would stress the maths portion of the processor more. And it has a fairly 'unadjustable' load... would that make a difference? Dunno. Usually experiments are easy (but boring) to do, but hard to properly design.

Oh well... I don't have a proper component level lab available to me any more, am I sorry? Sometimes.

 

[crow_smiling]

Take a look at this article and in particular Page 4.

 

[mviara]

Ah, finally some pictures and test setup! Thanks a lot, crow_smiling. I understand things much better now.

One point of confusion still remains. On page 4 of the article, it says:

Because the CPU only draws power through the 2x12V connection and nothing else does, this tells us the amount of power consumed by the CPU and the voltage regulators on the motherboard.

But earlier,

the 6/8pin aux socket does solely supply the cpu, but its not the only supply to the cpu.

So checking that connection may only give you some of the amps

Can someone shed more light on this?

Also, about the custom built shunt in the article, if I understand correctly, they cut the power cable and put the sensor in series, right?

Another question: I recently read about "Lights out management" software for servers. Some of them can supposedly monitor the power consumption over time. For example, the Apple XServe management software can report power statistics for each rail:

Power. Shows the current, voltage, and power draw, as well as a historical line graph for each supply rail. It also provides Uninterruptible Power Supply (UPS) information and status and monitors the redundant power supply option.

Has anyone here used such software? At what granularity (in time) is the power reported? Wouldn't this be an easy (and relatively accurate) way to measure the server CPU power?

 

[13thmonkey]

Ah, finally some pictures and test setup! Thanks a lot, crow_smiling. I understand things much better now.

One point of confusion still remains. On page 4 of the article, it says:
Because the CPU only draws power through the 2x12V connection and nothing else does, this tells us the amount of power consumed by the CPU and the voltage regulators on the motherboard.

But earlier,

the 6/8pin aux socket does solely supply the cpu, but its not the only supply to the cpu.

So checking that connection may only give you some of the amps

Can someone shed more light on this?



I've just skimmed the article, I am fairly sure (>95%) that the CPU does not draw power solely through the 2x12V aux supply, mainly because the ATX spec I believe has included these to increase the current available to the CPU over and above what was available from the std ATX connection. The voltages would be the same across this 12V as across the atx 12V so iff one dips the other would dip.

I'm ready and happy to stand corrected. I haven't dug out the ATX spec in a while, might go hunting for it in a minute.

Pin-outs for the ATX spec.

I'm not sure where all of these wires end up on the mobo but.. p36 shows the pinout for the main ATX connector, which has +12V1 still on it. Now unless there was a significant reroute when the aux connector came in (which is feasible) I'd say the atx connector still provides some power to the CPU.

I'm not sure how I'd find out more right now...

 

[GreenJelly]

I like it when my computers heat up my room in the dead of winter... Now thats power!

Id look long and hard at the intel website... they tend to provide this sort of information... but you got to look hard, its usually like finding a needle in a hay stack.

As far as REAL time power data, on a computer you are running... you can, but just dont... its not worth the time... You might spend the next 4 years in college learning the basics... hehe

It takes a VERY expensive scope ($20K+) ... One that will read amps (w, ohms, volts) billions of times per seconds, and it is WAY above my head to tell you how to do it if you did have the scope...

Becarefull... if you dive in too deep you might find yourself with a 4 year degree... or worst... you may be a PHD... then we would all have to call you DOCTOR! I would cry, the moon would glow like the sun, the seas would boil... and my carefully cut grass in my front yard would be ruined... YOU HEAR ME!!! STOP NOW BEFORE ITS TOO LATE!!!

 

[kwalker]

Hello. This seems a basic question. But I have not been able to find specific answers on Google and even on Tom's Hardware. Can anyone point me to an article that describes how the CPU (not the entire system) power can be measured?

I understand that this could be difficult to do, as one may need to insert a sense resistor before the CPU. Tom's Hardware must have a good experimental setup as they regularly report power numbers for various CPUs. I have not been able to find the article in which they report their measurement methodology - can anyone point me to it?

I have also been looking for boards that come equipped with such sense resistors. One such board is the Radisys Endura LS855 that uses a Pentium M processor. However, I need to use a dual-core processor for my experiments, which deal with dynamic thermal management for multi-core CPUs, so the Radisys board would not be suitable. I found this Intel article that uses a board called Jamison Canyon CRB that has a Core Duo is already instrumented with sense resistors for measuring CPU power. It seems like it is an Intel internal board, does anyone know more about this, or similarly instrumented boards?

First of all there are many variables to include in the measuring process.
To measure amperage you would need a shunt device to keep the circuit alive to make accurate measurements
The CPU has many voltage pins and is not at all active simultaiously.
You would need to isolate the circuit powering the CPU and stress the core or cores using software like prime 95 or many others available free.
Once you know a few variables you can calculate wattage using ohms law.
The trick of course is where do you use the shunt?
I ran across this link.
It may or may not suit your needs
http://wbta.us/Joomla/index.php?option=com_content&task=view&id=13&Itemid=1


If you’re an overclocker then this link might interest you
http://www.benchtest.com/calc.html

 

[mviara]

kwalker and greenjelly, thanks for your replies. I'm already a PhD student (can't get any worse, right No, no post doc plans), so I don't mind diving a little deep. Let me give an idea of why I'm doing this:

You know that chips are getting hotter, and it is impractical to design thermal solutions that target the worst case. My research deals with controlling the voltage/clock speed/throttle ratio of a processor while it is running to extract maximum performance in spite of thermal constraints. If this can be done reliably, then processor manufacturers will no longer have to underclock their chips - they could release chips with large worst-case to TDP ratios, and then let the thermal management software/firmware try to make use of high speeds cleverly in bursts.

There exist cycle-accurate simulators (Google Simplescalar, Wattch, Hotspot, and PTScalar) that could be used instead of an actual test rig. But there are many limitations - the biggest being simulation speed (few days to a couple of weeks per benchmark). Also, the simulators assume perfect knowledge of performance, power and temperature models, but our research intends to investigate issues when these are not that well known, as would be the case on real systems. Hence, we need an actual testbed.

From the earlier posts, it seems like what it takes is inserting a current sensor in between the 12VAUX cable and connecting this to a DAQ. I don't need very fine granular power measurements, as thermal time constants are larger (seconds). If so, I should be able to do this in a few weeks at most, certainly not 4 years ;-)

 

[GreenJelly]

I think what your talking about is handled differently in the core2duo processors. It sounds like you want to control voltage in a real time situation, based on performance... The Intel Core2Duo controls voltage in a real time situation based on what parts of the chip is beign used. It constantly turns on and off the portions of the CPU that are not beign used. This is done on the chip, and cannot be disabled as many have expressed in the past.

The process you request is beyond what 99.9% of us are able to do. I use to know a few guys who got their microchip engineering degree while at RIT... Many of them moved, and work at Intel. RIT has a great clean room... One of the best in the country.

This ME degree doesnt deals with some of what your looking for... But I would be MORE interested in talking to some computer engineers. They are the ones that build motherboards and other circuits controlling everything from cellphones to main computers.

From what I understand... The process of development for circuit design is this

Specifications Settings/Cost Analysis etc -> Circuit Requirements and solid state device requirements -> Circuit Design & Solid State Design -> Prototype Design -> Testing -> Process Engineering and Mass Production Design -> Testing -> Product Release and Software/Firmware design.

My Experiance with closed circuit processes was allot simpler
because the EE who I was working with using pre-existing products combined together. The outputs were well documented, and the communication speeds were on par with a Com 1 port.

Anyways we did a ProtoType -> software Design -> testing -> software design -> back to testing -> loop...

I also expect that certain motherboards and chipsets do what your talking about... Overclockers tend to disable these features because the MoBo manufacturers havent gotten them down right for higher then average settings.

I dont beleive ME, EE and CE individuals come to forums like this, because their issues and passions dont rely on Overclocking and system administration. In my case I have a passions, and a degree, that doesnt rely on the things generally posted on this forum. System building is mearly a hoby... At work I tell IT and Business men what I need, and then I start working... They get cost analyse the stuff and set it up. They generally buy systems from reliable corporate supliers who support their products at a different level.

I work on forums dealing with programming with the purpose of delivering information (MIS). I love to program for businesses, building good UI's, Pure and Strong Relational Databases, and strong OOP builds.

Unfortunately at this time I cannot program so I am here helping people with IT issues. I would go to the forums more suited to Engineers to get your information. Unfortunately I am not certain which forums exist.

These are the forums I use when I am up and running are...
1) postNuke (postnuke.org)
2) http://www.zend.com/ (php)
3) Microsoft.com (asp, sql, iis, and others)
4) Doom9.org (Hobby)
5) Oracle
6) Slashdot
7) Astalavista and other security based sites

I dont visit Tomshardware unless I am looking for build questions... I also havent had or seen any topics here regarding server design and builds. I think this is because we all go where are passions lead us...

The first thing you will definately need is to contact Intel for information and search their site. A scope may be needed so that you can actually read the voltages comming from each of the pins on the processors in a way that Is more compatable to what your looking for... I know DDR once used 1.5~0 voltage changes for each 1/0... This was 6 years ago, so my guess is that things changed. Also each pin on the cpu may be lower then 1.5! Generally everything Ive heard, binary signals are sent at under 2 volts, and they ATTEMPT to get to 0, though I have heard this is nearly impossible. These VERY low voltages, and the fast speed that the chip works makes things very hard to detect.

You would also want to use a real time simple opperating system. I hear there is a RT version of Linux. This will prevent you from receiving data from multiple threads, and allow you to control data by sending only certain commands to the chip.

I would also assume that a few pins deal specifically with voltage from the powersupply. How that voltage is used, is another matter. I would then expect that other pins use voltage drawn from other sources.

Your better off finding a good forum for circuit design and/or MicroEngineering... I dont have any recommendations to point you to... Good luck in your search...

Mike

P.S. Everything Ive seen of these forums are centered around individuals home/hobby computers. I havent seen any threads talking about network or server administration/setup

 

[mviara]

greenjellly, your reply makes sense. Perhaps this is not the right forum to ask for these specifics. I started here because I believed that surely, Tom's Hardware knows how to measure CPU power, and so should those who frequent its forumz. I did learn a lot from the earlier replies perhaps because I knew almost nothing to begin with. You're right about asking Intel - I will certainly try that. If anyone here thinks these questions can be better answered at other forums, please point me to them?

 

[mayouuu]

man :lol: , put it under full load with a heatsink filed with 100g of water xD, measure the time that it takes to the water to estart boiling and using fisics ( latent hot of water,delta T(100ºC-Initial temp( ºC).....) you can calculate aproximately your cpu heat output, it will no te so exact but try to minimize heat losts in some way and youll get a very aproximate value xDDDDD, im joking a bit......... xDDD