New Power Measurements

Some time ago (must have been over a year) I did some power measurements with a simple at-the-wall-socket power meter. For a short while they were on sale at electronics stores all over the place, as part of a "know your power use" campaign.

I’ve been sick for the past week (again) and one of the therapeutic things I did was rearrange my home office ; now the wall socket is easily accessible again and the servers are piled one on top of the other. I decided to do a little measurement and see what computation actually costs me. Bear in mind that this is totally unscientific, based on running a complete system through a power bar plugged into a meter and then watching the meter and guessing averages.

The system under test is an Intel i7-860 with 8GB RAM. Possibly the most difficult part of this test is actually putting the system under load. I’m hard-pressed to get more than, say, 25% load. One (single-threaded) simulation, one dual-core virtual machine running OpenIndiana and compiling KDE .. I tried to get kdesrc-build working quickly to see if that would load the machine too, but that was a bit of a disaster (aside from I’d forgotten that it would take an hour of just network activity to get the sources up-to-date). Something for another blog entry.

Anyway, here’s some brief power use information, for the complete system — that is, monitors and other peripherals that are switched on when I’m at my desk.

Samsung 206: 10W when "off", 39W when on, 10W when it goes to standby.
Iiyama E430: 10W when "off", 33W when on, 10W in standby.
CPU: 10W when "off", 64W on boot.
Complete desktop: 140-190W while logging in (after cold boot), 130W at an idle desktop, 170-200W when under load.

There’s two things I’m going to take away from this: at 200W total power draw for my desktop, that’s a ridiculous amount of computation power per watt. Second is that I’m going to switch off the power bar when I switch "off" the machine in the evenings, because 30W for standby is a bit much.

6 thoughts on “New Power Measurements

  1. When I had to do power consumption tests, I used cpuburn, which is specifically designed to burn the CPU, one instance per core. It overheats poorly cooled systems in a couple of minutes. The power consumption was much higher in my experience than a compilation-based job.

    Bad news is that it hasn’t seen any updates recently, so perhaps it doesn’t burn as much as it could newer processors. On early core2-s, it was amazing.

    If you want to put yourself in the worst case, reserve a core for I/O (say, grep -r “hello” /). In my experience, I/O power consumption was only significant on machines with many disks (say, 4-12), on desktops it’s usually irrelevant.

    OTOH, 130W idling is a lot for a desktop!!

    • @lfm: ah yes, that makes the system fan run much faster 🙂 each burn* command takes up around 12.5% CPU; up to six of them do take 100% of a (HT) core. Numbers seven and eight get less than 100% due to other things actually running such as Xorg.

      With lots of these guys running, power usage tops out around 220W total.

      Bear in mind that my “total” numbers (including the 130W idle) include the cost of two monitors which is around 70W. If I had speakers I’d leave them switched on, too, for the complete system power consumption picture.

    • @moltonel: both monitors report around 10W usage at the socket once I plug them in, so there’s clearly a big gap there. On the other hand, they are both more than 2 years old, so not subject to any recent efficiency regulation.

      • As all measurement hardware, there is a minimal power to allow detection. It seems you have not tried to measure with all hardware powered off : that will not show 30W, but 10W also.

        Anyway, yes to is a good idea to switchoff the plug when everything is down. By my side, I use a laptop, as it only uses 30W-60W powered up, and 2W in suspend-to-ram.

        A very diferent point : could you please tell me if the subtitles a friend has done to me in Belgian’s Flamand for the open-source movie shown in my website are also the good language in Nederlands? thanks.