Slashdot Mirror
Google Calls For Power Supply Design Changes
Raindance writes "The New York Times reports that Google is calling 'for a shift from multivoltage power supplies to a single 12-volt standard. Although voltage conversion would still take place on the PC motherboard, the simpler design of the new power supply would make it easier to achieve higher overall efficiencies ... The Google white paper argues that the opportunity for power savings is immense — by deploying the new power supplies in 100 million desktop PC's running eight hours a day, it will be possible to save 40 billion kilowatt-hours over three years, or more than $5 billion at California's energy rates.' This may have something to do with the electricity bill for Google's estimated 450,000 servers."
You've never seen a console port on a disk array, router, switch or UPS? That RJ-45 socket speaks RS-232 and will connect to your serial port with the right cable.
(Yes, some UPS's have USB)
Hands in my pocket
3.5" IDE disks still use the 12V rail (Dunno about SATA, but I would assume so, since the hardware is almost the same). Those cheap USB-powere drive cases are 2.5" (e.g. LAPTOP drive) cases.
n/m found it.
t e_paper.pdf
http://services.google.com/blog_resources/PSU_whi
Almost nothing, but that is irrelevant. A modern PC uses so much power that it would be plain stupid to try and deliver it at lower voltages. The power is stepped down close to the actual load, because otherwise you'd need much heavier wires or lose much power to heated cables. That kinda is the point of the proposal: Every PC already has the necessary regulators because there's simply no other sane way to deliver the kind of power that graphics cards and CPUs consume. So what's the point in keeping the power supply complicated when the main consumers in a PC use the 12V line anyway?
This is true, but Google is not throwing 7950's in their servers. These systems run with on-board video at best. Google has no need for a video card that can do anything more than text, as with all non-windows based servers. For that matter, after the first boot, there is no need for a video card at all.
There is no "I disagree" mod for a reason. Flamebait, Troll, and Overrated are not substitutes.
Modern CPUs run on core voltages of 1.5 v or less, depending on model. DDR RAM is 2.5V IIRC.
So you will have to convert most of your power from 5 V to something else. And if you have to re-convert anyway, 5V as intermediate voltage is not optimal. When converting to 5V, the voltage drop in the power diodes and in the wires to the mainboard eats a much higher proportion of the power than with 12V as intermediate voltage.
24V or even 48V would be even better. The auto industry is currenly starting to introduce 48V systems BTW.
C - the footgun of programming languages
RS-232
Sorry no. Modern rs232 circuits, if it's not already built into the UART, use a chip like max232 that runs off 5V and has a built-in charge pump to generate (close to) RS232 output voltages.
After reading Google's blog entry on the subject, I'm left puzzled by their call for a new standard with no further details, especially since it seems they're already using the technology. A power supply is simple enough, but I'd like to see what sort of strategy they're using for voltage conversion on their motherboards. What connectors are they using for power?
The funny thing is, this idea is relatively old, though AC was used instead of DC. Remember the Imsai 8080? The S-100 bus used an 18V AC supply, and each card had its own DC conversion and voltage regulator(s).
If you look at just routing 12 volts everywhere, you just would have to put the regulators in the hard drives, and CDROMS so they don't need 3.3 and 5 volts. Then what do you do about +5 Stanby that allows you to hibernate? Do you still need a stand by voltage? It isn't and easy answer and will take the whole industry to adopt it. Checkout formfactors.org for ATX and BTX specifications that Intel is pushing. What's also interesting is the 600 and 700, etc Watt power supplies just keep their 3.3 and 5 volts at around 30 amps max, but keep adding +12V1 +12V2 +12V3, etc.. Looks like the industry is already going to mostly 12 volts for distribution anyway. But don't you still need PS_ON, PowerOK, etc.. You're just trying to phase out the +5 and +3.3, and -12 which hardly any motherboards use these days, and maybe the +5 Standby, then it's going to happen eventually anyway. Most of the power is going on the 12 volt lines anyway, so having inefficient +3.3 and +5 isn't really a big deal. I've studied this for a while as my big hobby is computers in cars, I built a power supply called DSX12V that takes a 8-16 volt input and makes a solid 12v output that I got over 97% efficiency on. This is good for people sticking computers in cars or running them off banks of batteries for solar power applications etc.
no it does not: I=U/R, U gets lower, your personal R stays constant. Now if I is 50mA you're safe.
Google is proposing 12V between the server's internal power supply and the motherboard. Everything outside the server would still be 208V.
Their answer is the Petabox. It's a server setup designed to be "shipping-contained friendly", meaning they can build out a container stuffed with these racks, and have it operational on site with connections for power, cooling, and bandwidth. With this design, they can deploy a mirror of the Internet Archive anywhere that's willing to host it, without having to build a machine room or individual racks on site.
Capricorn Tech of San Francisco builds these machines and their site has more info.
--Pat
Video cards with a disk-drive-type power connector always use point-of-load switch-mode supplies to convert the +12V to whatever voltages are needed by the chips on the board. Nothing on the board uses the 12V directly, except maybe the fan (if that).
They use the disk-drive connector because:
Disk drive motors use 12V. Laptop drives (2.5" drives) use 5V exclusively, but standard desktop and server drives use 5V and 12V. SATA drives get 3.3V, 5V, and 12V. The VRMs that power your CPU and video card probably take their power off the 12V rail, as do many other components.
The reason you wouldn't want to power a machine off 5V is because you would need huge busses. Suppose you've got 40 svelte 1U servers in a rack, each drawing 100W. That's 4kW. Assuming that's a purely resistive load (hint: it's not), you'd need 800A at 5V for the whole rack. Are you familiar with the big connectors on car batteries? They're designed to pass less than half the 800A you'd need to run a rack off 5V, and your car battery only has to handle that for a few seconds while the engine is starting up; a rack would need to deal with that continuously. And that's for a pretty low-power rack.
Using 12V instead of 5V lets you get away with busses about 40% the size. Also, and probably more importantly, 12V DC is (IIRC - correct me if you're a PSU designer) easier to get efficiently than 5V DC. Once you split the 12V off into a few dozen servers, you can drop it down with small, fairly efficient CMOS regulators.
High-speed Road Trip (18.000KPH)
What google is proposing is supplying only 12 VDC to the mother board instead of the multiple power rails that go to todays power rails. This says nothing about how the supply of 12 VDC gets there or where it comes from. Frankly the idea is a good one though I have to wonder if 12 VDC is the optimal voltage for this.
As has already been mentioned Apple has computers that run fine off of 12 VDC and there are others that also work well with a single power supply rail. This is not a big deal and in fact is good modern engineering. Often it is referred to a point of load regulation and is happening on PC mother boards today! In fact many point of load regulators for the CPU on motherboards are running off 12 VDC. Many other odd voltage loads run off other supply rails. All google is really saying is to get rid of the multiple supply rails and standardize on 12 VDC and allow point of load regulation from there.
I hope every one grasps this now. The discussion of external facilities supplies don't fit in to this discussion at all.
Thanks
Dave
12V happens to be a sweet spot in terms of cost of the converter components as well as overall efficiency. Wire guage, voltage drop and capacitor size is significantly smaller than 5V or 3.3V primary supply. Think in terms of millions of units per month and compare the price of an NMOS FET and storage capacitor rated for 35V (safety margin in a 24V design) verses the cost of similar FETs and capacitors rated for 20V. In a synchronous buck design you can easily save $0.75 per converter section by using 12V rather than 24V and significantly increase conversion efficiency for free. Assuming a constant switch frequency the switching losses increase with the square of applied voltage, "I squared R" conduction losses in conductors will decrease with the square of current but the voltage dependant switching losses will dominate once the input voltage gets too high. For a given cost the overall converter efficiency is usually highest if your input voltage is relatively close to the output voltage. 12V to 3.3V conversion is significantly more efficient and less costly than 24V to 3.3V conversion.
Be careful - some network appliances actually use connectors that you'd never expect to do curious things. None of my PIXes (PIXen? PIXies?) have USB connectors...they have RJ-45 connectors, for which an adapter/cable exists that pins out to a DB9 connector. (I have made my own converter to allow for various configurations of this.) Dell switches have an SPF module that uses USB cables for stacking - USB A-A male cables to be precise. Those are most definitely /not/ for use with a computer. But really, I'd be quite surprised to find Cisco using USB - you'd practically have to have special drivers on the computer you connect with. Almost anything (save for a Mac) can run a serial console to DB9. Heck, I have a wonderful floppy - Serial Terminal Linux, which I can use to boot an old laptop right into minicom. Long live /dev/ttyS0 | COM1!!
The only things that natively use 12 Volts at a current high enough to be significant are the drives.
I think that misses the point however. Designing a power supply for higher output voltage, and switching regulators for higher input voltage, raises efficiency. (24 or 48 Voltage would likely be better yet, except for the need to come up with 12 Volts too)
It is unfortunate that the article (and the others that I could find) don't link to the white paper for some specifics.
Instead I'll have to base my comments more from my understanding of electronics.
Let's say that we start with the usual rectifier/filer off the power line feeding switching transistor(s) that pulse current into a high frequency transformer. That much is fairly basic and efficient if the transistors have a low on-state resistance, fast switching time (to minimize the power-burning partly-on interval during switching where the transistor has significant current flow through it and significant voltage across it), and the switching design is such that current stops before the transformer core saturates. We'll also assume that at the high input voltages involved the resistance losses of the transformer primary are small. The total power capacity of the supply is generally defined by this portion of it. How much can be delivered from individual outputs depends on the design of what follows.
Now we get to to output(s) from the transformer secondary, which is where the article indicates that having a single output, at 12 Volts, improves efficiency. Unfortunately we're given no reasons as to why.
1) By having a single output, the transformer secondary winding for that voltage can be of a heavy guage designed for the full rated power level. It avoids the problem of having to guess which outputs actually need to deliver most of the power.
2) If the power level is developed at a higher output voltage, the current is reduced making resistance losses smaller. That applies to resistance losses in the cables from the supply to the motherboard as well as resistance in the transformer secondary windings. Resistance losses increase with the square of the current.
3) There is a relatively fixed voltage drop (loss) in the rectifier(s) used to convert the transformer output to D.C. The higher the design output voltage is, the smaller a percentage that voltage drop becomes. Less current is required for a given power level at a higher voltage, so the loss in the rectifier(s) is reduced. These losses come closer to varying linearly with the current (over the normal operating range), especially when hefty rectifiers with low series resistance are used, something that is more likely in a well designed single-output power supply.
Is there something magical about 12 Volts? No. In fact designing for a higher output voltage would give even better efficiency, but then there would be the need to add electronics elsewhere to down-convert to make 12 Volts available.
If the portions of the machine using 12 Volts could tolerate the voltage variation, direct battery backup operation could be possible.
Since some modern CPUs may even have varying voltage requirements within a product family, and perhaps even unit to unit, there is a need for locally controlled voltage down-conversion on the motherboard. Done properly this can be pretty efficient. Generally the efficiency here is also higher starting with higher voltages. So if a CPU needs something near 2 Volts, it is better to produce that from a 12 Volt supply than from 5 Volts, for example.
The switching regulators on the motherboards can bring their own design issues. The ability to overclock some CPUs on some boards is limited by the available output current. Ability to recognize a new CPU type and generate a different voltage it needs is a more complex power-issue than we used to see. (This kind of thing could limit whether or not a Socket 775 motherboard that handled a Pentium D could handle a Core 2 Duo CPU for instance).
Actually, when I learned it, I was taught ~32V was the sweet spot between practicality, I^R losses, component size, and load type. Mind you, back then it was electrical / electromechanical loads (lights, motors, contactors, relays, etc) not electronic. And that was from a telco background (-48V), which made me wonder "well, why tell us that about 32V?" ;-)
But the point is, what the sweet spot is depends mostly upon the characteristics of the load - so it's wrong to come out with blanket statements like "12v happens to be a sweet spot in terms of cost of the converter components as well as overall efficiency". Yes, today, particularly with switchmode supplies and the actual maximum load V being 5v or less, it is. Tomorrow, when everything runs on 3.3v or less, it'll be closer to 5v~6v.
The other half of your argument only holds for certain types of power supplies too - but I'll give you a pass on that, seeing as you did explicitly state "synchronous buck" designs. It doesn't necessarily hold true, however, for other classes like linear, boost, buck-boost, etc. Your final assertation, however - that, for a given cost, the overall converter efficiency is usually highest if your input voltage is relatively close to the output voltage - is spot-on. Too far away from that, and the ol' V=I*R rule starts to bite you...
What part of "a well regulated militia" do you not understand?
There are different kinds of UPSes that do this in different ways. The two major types used for PCs are called "line interactive offline" and "dual conversion online". The first just passes the AC power straight through to the output. If the AC power coming in goes out of range, then it flips a switch internally (relay, contactor, thyristor, etc) to supply the power from an inverter driven by the battery. The second converts the AC coming in to DC all the time, and converts that DC back to AC for output. It then does the switching in DC, or parallels the DC with the battery directly. These variations are classified as "topology" by many manufacturers.
Both of these kinds can have inverters that produce square waves, pseudo-sine waves, or very nice clean sine waves. The dual conversion type can also isolate a poor power factor (the deviation of the current wavefrom from being a sine wave in sync with the voltage) of the PC power supply from the power source. A poor power factor means the product of the average current times the average voltage (apparent power) exceeds the actual real power (average of all the products of the voltage and current and each point in time) being used, which results in reduced efficiency and other problems.
now we need to go OSS in diesel cars