Should've overclocked CPU and GPU. Together and separately. Would be insteresting to see total system power consumption with i7 920 o/c to ~4Ghz
I bet CPU o/c would add another 40W or so, plus video card o/c, another 30-40W. And we're at 600 easy. Add another 2-3 HDDs (who has 1 HDD nowadays anyway), and that's 30W+ more. And what about after market cooling and extra fans? Or maybe water cooling? We could get to 700W under load. Granted, 800W PSU would do just fine. However, a 1KW PSU would be running under 40-50% load most of the time - sweet spot for power efficiency.

Congratulations, you did a beautiful job Mr. Artamonov.

I have never seen such a detailed and forthright look at the true challenge of measuring consumed power in a PC. Your decision to build a dedicated 8 channel current metering card is brave and brilliant. Being an electronics technician who sometimes works in the solar industry, I can deeply appreciate the difficulty in measuring multi-channel DC current systems.

But of course, it begs the question: When can we buy your current measurement card and software?

And perhaps a follow up question: What software / hardware development tools, and what PCB fab services did you choose for this project?

Of course, I fully understand if you don't want to make announcements just yet, but here's to hope. You have done an eloquent job of a gnarly problem that 99% of everyone else seem to simply sidestep.

Cheers

Great article. Can you talk a little about the test system's calibration, accuracy and repeatability so there is a baseline?

I noticed the Hall effect sensor data sheet shows a Total output error of 1.5% (typ)at TA = 25°C and then the graph for mean total output error over temp shows about 3% at 25C and then falling off to under 2% at high temp... Add in the uC's ADC nonlinearity and offset/gain error of the analog gain stage and you might see 2-4% offsets/error in the results. Again, not a big deal, but it'd be nice to know and might just be a one-line statement with any future articles.

Thanks again.

It's definitely an interesting article, device and software, however the assumptions that people are purchasing far greater power supplies (and more expensive) than they require, as well as manufactures "declare overstated specifications", I don't think is generally the case.
Most people I know get the general power requirements from the product manufacturers.
min req. from EVGA:
GTX 285 - 550w
GTX 295 - 680w.
The listed power requirements by the card producers fall right in line with the measurements that you Mr. Artamonov ultimately produce; allowing for a few extra drives, some overclocking and my USB powered coffee plate.
Manufacturer listed p/s req. seem to be about right.
btw, I believe that a real gaming computer running Crysis w DX10 at 1920 x 1200 will peak power draw greater than FurMark as Crysis will load both the CPU @ GPU quite intensely at times. It would be helpful if you gave some specific power draw data for a popular demand requirement such as this.
Thanks for your work, respectfully,

Oleg, great article! Finally concrete results about power consumption.

I'm intrested, did you consider power efficiency curves? What about the idea, that efficiency is greatest at the half of maximum power? Although PC rarely works at peak long time, a bit of reserve power is good for general efficiency when PC is under normal load.

the software 'OCCT' and the tab 'power supply' does a better job of maxing out the psu power usage.
(by about 30-40 watt)

I before I used the perlin noise benchmark in 3dmark06, but the occt test is better at consuming power.

(but nothing works better than prime95 for finding an unstable overclocked computer..)

The article measures some real-world systems that are extremely inefficient in performance/power and while a lot of real-world systems are like it would be better to have shown systems that are efficient at various levels of performance. For a person building a system, there is little reason not to choose the efficient mainstream components.

For example consider these systems:http://www.silentpcreview.com/article943-page5.html
High performance office/home/htpc/very light gaming systems more powerful for average uses than the article's home system and much more powerful than the office system, but with power consumptions of 35W idle and 65W full load. This is measured at wall including PSU ineffiency, so PSU output needed is sub 65W. A 3.5" drive should be used instead but will only increase power by a few watts.

I second the suggestion to try out some really high overclocks and then measure the power usage. I have always claimed that 750W is enough for almost everyone but people never listen to me.

I am personally running a 750W PC Power & Cooling PSU with the following hardware and overclocks:
QX9650 @ 3.825 ghz and 1700 fsb (will be pushed to 4.05 ghz and 1800 fsb this weekend)
4 gb ddr 1066 5-5-5-15 ram
2 x raptors
2 x 285 ocx's in sli
2 x optical drives
x-fi elite pro

A very interesting article, I've learned quite a lot too, thank you.

An exceptional article, that any computer builder should read.

Are you considering rail power limits? I see about 40A on 12V in your last table. How many 500W power supplies do 40A on a 12 V rail? Zero? 650W PS? Some do, but how many do it on a single rail, or the one that is going to the gfx/GPUs? Not many.

40th floor, I don't see why that should impose a problem. Even on split rail PSUs (say one with several 20A rails), the power for the graphics card will be divided across rails, it won't be limited to coming off one single one.

Hey Oleg, nice job and great detail ... I just built one channel similar to your setup and used it to measure the 12 VDC on the CPU of an old sysetm (I used the ACS713 that you used). Even uncalibrated, I measured the CPU power of a QX6700 within a Watt or two of your CPU measurements (published in prior articles).

I am finishing up gathering parts to finish out a full 8 channel setup and am excited to see the results.

What is your background? -- Xbit's power data and info always seems to be spot on and your methods a cut above any other site.

Thanks again for the info, nicely documented methods (i.e. reproducible), and the wonderful idea -- beats a shunt (hands down) for data collecting and ease.

Jack