I’ve just read an interesting post on TomsHardware.com about a solar powered PC [1]. It describes all the steps involved in creating a modern high-performance low-power computer.
They have a lot of interesting information. One surprising fact (from page 3) is that the PSUs tested (both for AC and DC input) were more efficient when idle (I expected the greatest efficiency to be when under load).
An AMD processor was chosen due in large part to the fact that chipsets in suitable motherboards used less power. For the CPU itself Intel had a competitive offering but no matching motherboard was power efficient enough (from page 7).
Page 8 documents how using a cooling fan (instead of passive cooling) reduced the power requirements of the CPU to such a degree that it always saved power use overall. Why do CPUs take less power when they are cooler?
Page 9 mentions that a small passively cooled video card can draw 88.5W when idle! That sucks pretty badly, it seems that having a video controller integrated with the motherboard is the way to go if you want to save power.
It’s interesting to note how much energy can be used by RAM. Page 13 shows that the difference between 2*1G and 2*512M can be as much as 3.4W and that the difference between different brands of RAM for the 2*1G can make as much as 1.2W difference. Their final system drew 61W when idle, my latest 64bit system takes 52W when idle [2] (which compares to the 38W of their system without a monitor), so we are talking about 9% of system power being saved by using less RAM or 3% being saved by using a different brand of RAM.
The summary of hard drive power use on page 14 is interesting, the fact that 2.5 inch laptop disks use less power than 3.5 inch desktop disks is hardly surprising, but the difference when idle is very surprising (apparently one of the 3.5 inch disks spends 8W on turbulence and friction in the bearings). It’s unfortunate that they didn’t compare any of the server-class 2.5 inch disks, it was about 6 months before the article was written that HP announced that in future they would cease shipping 3.5 inch disks and only use 2.5 inch disks (I wonder if this is related to all HP’s recent work on server cooling). Rumor has it that many server class 3.5 inch disks have platters that would fit into a 2.5 inch case because at high rotational speeds a larger diameter platter would not be strong enough.
The information on DVD power use on page 15 is quite shocking. From now on when I install machines as servers which don’t have a need for for a CD-ROM drive I’ll remove the drive prior to deployment. Even if it saves only 0.47W then it’s still worth doing on a machine which uses less than 40W! An additional benefit of this is that it might speed up the boot process as the system won’t need to check for a bootable CD.
It’s unfortunate that most computer parts don’t have published documentation on how much power they draw. Even if you don’t want to run on solar power there are still significant benefits to saving electricity (including reducing the noise from cooling fans and heat problems in summer). If technical data was published then people could make informed decisions about which parts to buy.
Update: Changed the percentage savings for different types of RAM to be based on the system power use without the monitor. I’m most interested in saving power for servers and for idle desktops (running a desktop machine 24*7 is pretty common) so most of the time the monitor will be turned off.
It’s interesting to note that they power their system uses is about the same as a P3 system and could be less if they used a different hard drive.
There are several reasons why CPUs take less powerful when they are cooler. For an idle CPU, the main reason is that leakage current increases with temperature. For an active CPU, the main reason is that resistance increases with temperature (since delay is roughly a linear function of RC, higher temperature can increase the chance that operations can’t complete within the clock cycle, which is why overclockers chill their CPUs, sometimes with liquid N2).
Notice that there’s positive feedback here: if a CPU gets hot, these effects will make it even hotter, until thermal runaway results (modern CPUs will shut themselves off when they get too hot for this reason).
“Why do CPUs take less power when they are cooler?”
Lower temperature == lower electrical resistance. Heard of superconducting?
http://en.wikipedia.org/wiki/Superconductivity
Also, the most power-expensive operation a CPU can do is to drive signals off-chip (mainly to talk to RAM). Cache misses don’t just cost time, they burn power.
Joe: I guess it must be leakage. If the system is working correctly (IE the resistance is not great enough to prevent it functioning) then resistance could be expected to decrease power use.
Interesting point about talking to RAM. What we need is a module with CPU, motherboard logic, video, ethernet, and 1G of RAM on the same module. That would be quite achievable, such chips from a P3 class machine (apart from RAM) would probably use less transistors than some of the modern multi-core CPUs.
some guy: I don’t think that superconductivity is relevant to this discussion as it’s not used in CPUs at this time.
The reason *I* don’t publish anything like that on zareason.de is that customers would start measuring with their own devices – and none of these measurements are really comparable.
If I would test the same machine like Tom’s Hardware or the German c’t magazine, I would end up with totally different numbers.
That could cost vendors and consumers more than it would help, at least here in Germoney ;-)
cheers,
wjl
Just to make a note, their system uses 61W when idle, and yours uses 52W. But they’ve figured the monitor power in the number, and you haven’t.
You might find this interesting:
http://www.docunext.com/blog/2007/08/16/solar-powered-dectop/
Working at a hardware vendor, I know that OEMs are very interested in power requirements. You should find power consumption on data sheets. However, you might have to look at data sheets for individual chips rather than boards.
wjl: If you tested the same machine you would not end up with totally different numbers. The difference between a machine with 1G of RAM and a machine with 4G would probably be lost in the noise, but the difference between a P3 and a P4 (or other similar differences) would be obvious.
vedran: Correct. I’ve updated this post to be based on the system without monitor.