Slashdot Mirror
Google Demands Higher Chip Temps From Intel
JagsLive writes "When purchasing server processors directly from Intel, Google has insisted on a guarantee that the chips can operate at temperatures five degrees centigrade higher than their standard qualification, according to a former Google employee. This allowed the search giant to maintain higher temperatures within its data centers, the ex-employee says, and save millions of dollars each year in cooling costs."
Uhhhh. Wouldn't making chips a bit more efficient be better, as opposed to making them "less likely to burn out at higher temps"
Seems that google's not really thinking green in this case (despite the pretension to do so in others), unless they plan on making use of the datacenter heat elsewhere.
If you don't have the clout of a Google-sized organization to buy higher-rated chips from Intel, I wonder if you can basically achieve the same thing by underclocking. An underclocked chip will run cooler, but I don't know if it'll run more stably at higher temps, although I think it would.
Does anyone have any experience with doing this?
I think it'd be interesting to see whether the cost savings in power and cooling is offset by the cost of the performance losses.
You see? You see? Your stupid minds! Stupid! Stupid!
Under-clocking them a bit can't be that hard to do.
WI'm surprised Google isn't considering moving some of its data centres to Arctic locations where you get cool temperatures year-round.
There is no reason to be surprised. It is cheaper to not move the data center to where it is colder and just make all upgraded hardware use the new chips. Google's budget calls for hardware upgrades already. Upgrading to CPUs with higher temp tolerances would mean they pay the same $X-thousand for the box they would anyway and simply turn the thermostat up.
A net savings.
Two words: "Free Cooling"
The greater the difference between your data centre's output air temperature and whatever passive external cooling system you are pumping it though, the more heat you can dump at zero cost. That's monetary cost as well as the cost to the environment through the energy "wasted" in HVAC systems and the like. Google has a PUE (Power Usage Effectiveness; the ratio of power input to power acutally used for powering production systems) approaching 1.2 vs typical values of around 2.5-3.0 - Microsoft is around 2.75 as I recall - so they are clearly doing something right.
UNIX? They're not even circumcised! Savages!
You know, they wouldn't have to go that far. I'm in Colorado, if they put a data center in one of the higher mountain towns I imagine they could significantly reduce their costs.
I guess the other thing they could look at is using a heat exchanger and use that excess heat for hot water heating or something.
One should not theorize before one has data. -Sherlock Holmes-
Most of the power supply systems for my servers, which are HP G3-5 systems of various U sizes, tend to waste more power as temperature goes up.
This has nothing to do with CPU's though. It is the power supplies on the machines. As temperature goes up, efficiency goes down. At around 80 degrees I noticed a significant larger draw on the power supply with my amp meter.
I had a gaming system with two ATI 4870's and the 800 Watt power supply would crash my machine if I did not run the air conditioner and keep the room at 70 degrees after some fairly long Supreme Commander runs.
I noticed that the amperage would go up, and the power output would go down as temperature would go up.
I have not conducted any experiments in a lab setting with this stuff, but from experience, jacking the temperature up usually makes power supplies work harder and makes them less efficient.
-gc
Got Geometrodynamics? Awe, too hard to figure out? Too bad.
It also wouldn't surprise me if Google were willing to offer something of a testbed setup. A while back, they put out that report on HDD reliability and its influences, so they are obviously watching that sort of thing. And, since their historical style has been very much about redundant masses of commodity gear, they can theoretically tolerate slightly higher failure rates if those lower costs in other ways.
I suspect that, with negotiation to set the correct balance of pricing, warranty, access to handpicked chips, etc. both Intel and Google could easily benefit from an arrangement where Google gets to play with slightly experimental stuff, like higher temperature processors, and Intel gets field reliability data.
This happens with resistors, too. If you want one within 5% of the nominal ohmage, you pay more. If you want want one within 10%, you pay less, but you'll find that they're all either about 10% low or 10% high, with a 'notch' in the center of the distribution. Same production process used, but they skim off the 'good ones' and charge more for them.
PHEM - party like it's 1997-2003!
Wouldn't Intel run into physical limitations that simply don't allow chips to run at that low a temperature? I'm surprised Google isn't considering moving some of its data centres to Arctic locations where you get cool temperatures year-round. We've seen reports of appealing places like that on Slashdot before. (Of course, that would just be a short-term fix before we move the Earth to a farther orbit around the sun to avoid suffocating in our own waste heat like the Puppeteers in Niven's Ringworld ).
I doubt anything physical is being done. Intel is very conservative in setting maximum operating temperatures. They're probably just promising Google that they'll cover those operated 5 C hotter under their warranty. If anything is actually being done to the hardware it's probably just altering the temp at which throttling occurs.
"I zero-index my hamsters" - Willtor (147206)
Legend has it that the Celeron processor began its life as a way for Intel to make money off of the Pentiums that didn't pass quality control. If they sell the low performing processors at a discount, why shouldn't they sell the over performing ones at a premium?
One of our competitors trademarked the term "hypothesis". From now on, we will call them "boneheaded ideas".
It's going to be far cheaper to build radiator fins extending into space than to move the earth's orbit, barring some innovative invention in the orbit moving department. Also, orbit moving has the downside of reducing the solar flux, which will be bad for our solar energy efforts.
"Who is the Journal of Quantum Physics going to believe?" --Stephen Hawking
Hard disks. In fact, I am typically far more concerned with long-term issues with my data than with the computing itself. Not to mention, the CPU is NOT the only chip that can suffer from heat issues.
Iceland. It has cheap geothermal energy. And that's energy that's going to heat the Earth, anyway, similar to solar. They just need some big pipes between there are North America and Europe.
now we need to go OSS in diesel cars
Best isn't a term you use in test programs. At a certain temperature a chip will run at a certain speed or it won't.
All that's going on here is that Intel will be altering its binning process to separate out the chips that are capable of running at 75 degrees from those that can't.
Not really any different at all from the current process of binning based on speed grade. All it'll be is a different set of parameters in the test structures will cause the chip to go into a different bin.
Now what interests me is that if google are guaranteeing to run these chips at a minimum temperature and therefore could increase the yield by accepting chips that would otherwise have been a failure because they couldn't run at a cold temperature...
"The weirdest thing about a mind, is that every answer that you find, is the basis of a brand new cliche" -
Google's chips will be running full throttle/full temp 24/7
Is there any documentation for this?
I seriously find it hard to believe that Google has every processor they own running 24/7 at 100% utilization. Other than the computation problems like SETI and protein folding, most problems are I/O bound, and I would think that the stuff Google does would involve a lot of I/O.
You are correct, or it may not cook correctly at all at the lower temperature even with more time. For example, if you want to boil rice in Guatemala City, you have to sautee it first.
Damn Slashdot's broken Unicode support. There should be an accent in there.
$x='S24;r)>63/* h@<5+oZ)32"5cz';$me='phroggy'x$];
$x=~y+ -xz+\0-Tx+;print$_^chop$me for split'',$x;