Power Consumption of computing equipment

These measurements were made with a kill-a-watt.

measurements

machine specs off no load CPU burn disk read charging battery
crusty Pentium I 90MHz, 80MB RAM, 1G PATA disk ? 41W 43W 46W N/A
splat alix AMD Geode 500MHz, 256MB RAM, 2G CF ? 5W 7W ?? N/A
squeak thinkpad X30, 12" screen, P3 1.2GHz, 512MB RAM, 40GB PATA laptop HD ? 17W 31W 24.5W bat1: 58W;35W;27W;22W
bat2: 57W;48W;35W;25W
atlas dual PIII 1GHz, 768MB RAM, 2x200G PATA (RAID 1) 5W 78W 100W 87W (91W md1) N/A
clam 1st Gen iBook, PowerPC 300MHz, 92MB RAM, 3GB 2.5" PATA disk 2W 15W 17W 18W N/A (battery dead)
squash AMD K6 350MHz 384MB RAM, 6GB PATA disk 1W 30W 49W 44W N/A
echo gateway solo, 12" screen, PII 233MHz, 192MB RAM, 4G CF ? 19W (12W screen off) 25W 24W N/A
pip ASUS eeePC 900, 8.7" screen, CeleronM 900MHz, 1G RAM, 16G SSD (wireless adds ~2W to these numbers) 1W 13W 16W 14W 35W
smoke BioStar M7VKL Athlon 1GHz, 384MB RAM, 4GB 3.5" PATA disk 4W 76W 95W 78W N/A
cur FoxConn A78AX, Athlon X2 4450e, 4GB RAM, 2x500GB 3.5" SATA disks 2W 45W 83W (while syncing RAID) ?? N/A
loki Dell Dimension 4500S, 2GHz P4, 1GB RAM, 80GB PATA disk 1W 33W 69W 45W N/A

  1. Power Consumption of computing equipment
    1. measurements
    2. costs/tradeoffs
    3. strategies currently measured
      1. off
      2. no load
      3. CPU burn
      4. disk read
      5. charging battery
    4. other strategies
      1. networking
        1. wireless networking
      2. bluetooth
      3. USB devices
      4. serial console
      5. optical media
      6. random disk access
      7. disk write
      8. video
      9. memory activity
      10. sleeping
    5. units: watts vs. volt-amps
    6. Caveats
    7. other notes

costs/tradeoffs

Replacing crusty with splat, above, saves me at least 34W. that's around 300 kWh/year. Since my electric utility charges ~$0.11/kWh, i'll save ~$30/year for this one replacement. While this doesn't exactly meet the costs of the new machine, it is still pretty significant.

In addition to the financial costs, of course, there is also the lower environmental impact to consider (must research the impact of creating/shipping the new machine), the reduced acoustic noise, and the reduced heat output.

strategies currently measured

off

How much does the machine draw after the machine has been shutdown and the power turned off. The standard "shutdown" command for linux is: 

shutdown -h now

Most modern systems include power management systems like apm or acpi that will halt the kernel and power down the machine when issued the shutdown command. This is the state we intend to measure in this column.

Many older machines, though, do not have power management systems and will only halt the kernel when given the shutdown command, leaving the machine in a powered up state which can draw as much power as a no load state. For these types of machines, I think an appropriate way to measure power consumption of this state would be to issue a shutdown, then actually turn off (or unplug) the power, then return power without actually booting the machine. For machines where this is not possible, then this column is not relevant and an "N/A" is used.

no load

let a stock linux kernel boot (kernels from lenny and etch at the moment), make sure that vmstat reports mostly idle CPU cycles.

CPU burn

force the kernel to chew up as much of the CPU as possible. on single CPU machines, just ask the kernel to generate pseudorandom numbers and throw them away: 

dd if=/dev/urandom of=/dev/null

If you've got a multi-CPI (or dual-core) machine, you might need to run one instance of this for each core you have available to make vmstat claim 100% load.

  • This won't exercise the FPU on any machine; if the FPU consumes significantly more power than the ALU for the specific core design, this might not be as high as you could push it.

disk read

pull data straight from the raw disk device as fast as possible, and send it to the bit bucket. This doesn't affect the data on the disk at all, other than forcing a series of linear reads: 

dd if=/dev/hda of=/dev/null
  • does the kernel's disk cache have an effect on this?

charging battery

battery should be completely drained initially (0%).

measurements should be at no load, and then immediately when it starts charging, 50% full, 75% full, 90% full.

If there are multiple batteries, indicate each battery on its own line.

other strategies

There are obviously other interesting things that could affect power measurements. These would be worth measuring at some point, for the devices which support them.

networking

would sending a lot of traffic over the network affect power consumption?

wireless networking

bluetooth

USB devices

  • audio devices?
  • USB serial lines

serial console

If you send huge amounts of data over the serial console, does it affect power consumption?

optical media

Reading CDs and DVDs takes power, both for the optics and for the mechanics of getting the disk spinning. Do "music" CDs take more power to read than "data" CDs?

Writing CDs and DVDs takes power as well. Does writing at faster speeds consume more power?

random disk access

heavily random disk activity (e.g. swapping) -- straight dd is linear access, which might be different, since seeks probably cost something other than just active disks.

disk write

Warning: this is extremely destructive! Don't do this unless you understand what it will do to the machine you run it on.

Best to select a specific partition that you know you can trash: 

dd if=/dev/zero of=/dev/hda7

video

How do we get massive video terminal activity (random pixel flipping)

memory activity

i've heard reports that amount of memory present is an indicator of power consumed. Is active use of the RAM required, or is it just the cost of refreshing the DRAM? Could we use memtest86+ for this test?

sleeping

if the machine has a "suspend to RAM" state, how much power does it draw in that state?

units: watts vs. volt-amps

There is often confusion over which units to use when measuring power consumption. The two most common measures of power are the watt and the volt-amp. In AC circuits, the power is a complex number with both real and imaginary components. The real (in-phase) part of the power, P, measured in watts, comes from resistive circuit elements that actually dissipate power and consume energy (ie. resistors). The complex (quad-phase) part of the power, Q, comes from reactive circuit elements that just store energy to be released later in the cycle and therefore consume no energy (ie. capacitors and inductors). The "apparent" power is the rms sum of the two and is measured in volt-amps:

S2 = P2 + Q2

Volt-amps are calculated by taking the product of the rms voltage and rms current delivered to a circuit.

The ratio of real to apparent power is the power factor:

pf = P/S,

which is always a number between 0 and 1. The conversion between watts and volt-amps is then:

volt-amps = watts / pf.

The distinction between watts and volt-amps becomes important for highly reactive circuits that have low power factors, ie. computers. A computer with a power factor of .5 might consume 10 watts of real power, but would require 20 volt-amps of apparent power. Watts measures actual energy dissipated in the circuit, and is therefore what is billed by power companies. However it is volt-amps, which is always equal to or greater than watts, that must actually be delivered to the system. This is why power delivery systems (ie. UPS's) are rated in terms of volt-amps.

Caveats

  • Some mainboards have BIOS settings that can significantly affect the power consumption. For example, a setting to vary the speed of the CPU fan with the internal temperature can significantly (~10W) reduce the power consumption of a machine when it is cool enough to spin down the fan. Review your BIOS for settings that might be relevant.

other notes

  • powertop -- tool to find out what is using power on a Linux computer, with advice on things you can do to reduce power consumption