First Actual CPU Energy Use Statistics Published

BBCWatcher writes "CNN is reporting that the U.S. Environmental Protection Agency in August asked server manufacturers to develop 'miles per gallon' ratings for their equipment that would provide accurate assessments of energy efficiency. IBM says it is now providing 'typical usage ratings' for its line of z9 mainframe computers, in addition to previously available maximum power ratings. More than 1,000 z9s around the world started reporting (with the owners' permission) on May 11th their actual installed power and cooling demands, so IBM can publish statistics such as how much energy is required to turn on an additional processor to run multiple Linux virtual servers. The answer? About 20 total watts. 'Over time every vendor is going to be asked to provide typical energy use numbers for their equipment. It's what the EPA wants, and this allows us to move beyond simple performance benchmarking to energy benchmarking.'"

Comments

damn lies

[Bob-taro]

Oh, but there are so many ways to play with statistics. Hey, is the the first post?

Re:damn lies

[Bob-taro]

Ha! I confess, I cut that post short to try to get the first post (I'd never gotten one before!)

Anyway, the problem with trying to get some "miles per gallon" efficiency rating on computers is defining the "miles". For example, if computer A is 2 times faster and uses 1.5 times the energy compared to computer B at full load, and both computers are run at full load 8 hrs a day (doing some serious number crunching), which computer is more efficient? A is using more power, but is doing twice the amount of "work" of B. So do you measure straight Watts? Watts / MFLOPS? If you use MFLOPS, how do you account for differences in architecture?

Re:damn lies

[Colin Smith]

Oh, but there are so many ways to play with statistics All the more reason to understand how they work.

Re:damn lies

[vtscott]

I agree. While this could be a useful measure, companies will find ways to game the system like they always do. The problem is, if they just publish hard numbers and exact specs they will be difficult to interpret. When they publish these "more useful" stats, they'll just do everything they can to get the biggest (or smallest) numbers while sacrificing everything else.

For example... I was recently shopping for home theater projectors and was doing a lot of comparisons between brands. The two most important things I was looking for were picture quality and brightness. Both have standard metrics such as contrast ratio and lumens. The problem is that the manufacturers know they're being judged on these two numbers, so they just play with the numbers until they get them as high as possible. Sure, you can get theoretically get a picture with the advertised contrast ratio, but the brightness will be terrible. And yes, you can get a picture as bright as advertised, but it will look like crap. The numbers are supposed to give you an indication of the quality of the projector, but instead it's just a dick measuring contest between manufacturers and most would likely sacrifice overall quality just to up their numbers.

I would expect the same will go for CPU manufacturers in this case. Processors will be advertised at XGHz and Ywatts*, but you definitely won't see both X and Y true at the same time.

*Y watts based on our definition of "typical use" while running at a much slower clock speed than X

Re:damn lies

[dgatwood]

While you're at it, account for a more typical load of 10%, account for periods overnight when the machine may be completely idle (for a US-centric web site, for example), etc.

First, everyone has to agree on a series of load metrics which when combined are sufficient to get a good approximation of performance under various types of load (disk-heavy load, CPU-heavy load, etc.). Only then can we really answer the question of how much power a server is going to waste.... :-)

Power consumption since mid-80's?

[securityfolk]

I would be interested in seeing statistics from the 80's on, to show how much power an "average" home PC or business workstation consumes. Over the years I've heard things like "it takes more energy to power it on than to just let it run", "it uses as much energy as a blowdryer", and "it uses as much energy as a lightbulb". Also, doesn't it take LOTS more energy to continually refresh RAM than it does to enable a processor? Think we could break usage down by component in these stats?

Cheers, Securityfolk

The first one though...

[pigiron]

I'm sure it's true that adding additional virtual servers is only 20 watts. But that first one is a real whopper!

Nice to get a watt/CPU

[Lookin4Trouble]

How about getting a realistic number for BTUs of cooling per HDD/stick of RAM/Processor? my 31 year old Liebert is dying, and the time has come to go to in-row rack-standing AC, but I don't know whether to stick to 2x10Ton or if I need to go for a 3x10 (underfloor in a small datacenter - 30 racks, 250ish nodes). I realize manufacturers have whitepapers out on how much cooling is recommended, but those numbers lie like dogs. "Typical installation: 1 processor, 1 stick of ram, 1 HDD, 1 Power supply" - typical config for my cluster is 4 processors, 8 sticks of RAM, and 2 HDDs on dual power supplies... anyone know where I could get this type of info besides Gartner or the like subscription $ervice$ (yep, they get you coming AND going)

Re:Nice to get a watt/CPU

[An dochasac]

While it's a baby step in the right direction, Watts alone as a "benchmark" is meaningless as is Watts/CPU. The VIC-20 likely beat the Z9 back in 1980.

If IBM is serious about server energy consumption, they should publish statistics using the SWaP (Space Watts and Performance) benchmark Sun has been promoting for several years or even "MFLOPS/Watt" or "Page serves/second/Watt" If the Z9 can handle a typical highly threaded webserver load with fewer watts than something like Sun's T2000 Niagara while providing identical performance, IBM shouldn't be afraid to prove it.

Until then, I'll assume it's just another useless benchmark configured specifically to make IBM's products look better than its competitors.

energy efficiency has been tackled already

[QuantumRiff]

By these guys back in 2000. The potato powered web server.. We could help our farmers, and power our data centers with beuwolf clusters of potatoes!

Re:energy efficiency has been tackled already

[veganboyjosh]

...if only the sysadmins weren't constantly dismantling the power supplies, dropping them in hot oil, and eating them, that might work!

Re:energy efficiency has been tackled already

[argiedot]

So, the internet is a series of tubers then?

Reason To Buy A CPU

[bostons1337]

Just curious what you guys think about how this relates to buying a CPU. Do you think individuals and companies are going to take a big look at the CPU Energy Use when deciding on buying CPUs? I personally don't think it will become a deciding factor, like processor speed or L1 and L2 cache size, but I think it definetly helps in making a decision.

Re:Reason To Buy A CPU

[foobsr]

"The plan includes new products and services for IBM and its clients to sharply reduce data center energy consumption, transforming the world's business and public technology infrastructures into "green" data centers.

The savings are substantial -- for an average 25,000 square foot data center, clients should be able to achieve 42 percent energy savings. Based on the energy mix in the US, this savings equates to 7,439 tons of carbon emissions saved per year."

http://www-03.ibm.com/press/us/en/pressrelease/21524.wss (emphasis mine :)

I think otherwise.

CC.

Re:Reason To Buy A CPU

[bmajik]

My newegg order just showed up last nite. I wanted a machine that was as silent as possible, so I got an AMD BE2350 (the 45W TDP dual core Athlon @ 2.1ghz), an MSI k9 platinum (heatpipe cooling for the chipset), and a Gigabyte "silent-pipe" 8600 GT card.

For a power supply i got a seasonic 330w S12 (variable speed ballbearing fan).

My computer is entirely fanless except for the stock AMD CPU fan and the Seasonic power supply fan. There's not even a case fan. System and CPU temps seem to be stable around 40C.

My vista "index" is 5.0, with the 5.0 being the lowest number and coming from the CPU.

I wanted a really quiet machine. That meant eliminating fans. That meant buying energy efficient parts (the CPU and the Seasonic PS are both spendier than equivalent parts that don't stick to a tighter energy budget). But the machine _is_ quiet. I've got a kill-a-watt at home that I haven't tried out yet but I hope to see less than 100w of consumption. My old socket 754 machine is 5w sleep, ~100w booted but idle.

I'm also going to be consolidating my "always-on" applications (file serving, possibly BT) onto a Windows home server machine so that i can have my other boxes power-save as much as possible without any real service interruption. Having a few songs here, a few videos there, etc means that I can't keep the majority of machines sleeping the majority of the time (WOL is pretty spotty IMO.. if you configure WOL such that a machine "can" wake, it usually will stay awake from other network noise)

One of the other things i bought with this order was a new UPS. Sticking to a smaller power budget has other interesting effects -- like you can get away with a smaller (and cheaper) UPS to get the same amount of uptime.

Re:Reason To Buy A CPU

[Spoke]

One of the biggest power draws these days are graphics cards. Often, graphic cards will draw as much power as the rest of the system. It's typical for the cheapest graphics card to add 10 watts to the power draw of the system when idle, with 25-50+ being common for medium to high end cards. Unfortunately, current graphics cards don't do much in the way of reducing power draw when idle.

If you don't do any serious gaming, sticking with the onboard graphics will often reduce power draw significantly. If your mobo doesn't have onboard graphics, picking an inexpensive fanless graphics card will draw the least power.

If you were using onboard graphics, I would expect your system would idle around 55w (+-5w or so). Peak power draw would be less than 100w. With the GPU you're using, I'd guess that it adds add 10-20w at idle and another 50w at peak. It'd be interesting to see what the actual numbers are.

Something people often forget is that a good PSU with active power correction will also significantly reduce the apparently load on a UPS (as well as the grid if you don't have a UPS), not to mention that PSUs with APC are normally significantly more efficient. For example, if your system draws 100w but your PSU has a power factor of .5, you are actually pushing twice as much current through the AC line as a system drawing 100w but a power factor of 1. This ends up doubling the load on your UPS if you have one.

These days it's fairly easy to build a system which idles below 50w as long as you're informed. A bit more research will get you something in the 30-35w idle range if not lowre. I do wonder what you had in your old Socket 754 machine which caused it to idle at 100w. I suspect it had an inefficient PSU and a mid-high end graphics card or wasn't using Cool'n'Quiet. All recent AMD systems I've seen which support Cool'n'Quiet idle at 60w or less unless you have a power sucking GPU.

How big a fraction?

[columbus]

This looks like a positive development.

It seems that the computer industry on the whole has become more concerned with energy efficiency over the last few years. I'm glad to see it. As a discipline, computer science is always looking for ways to eke out more efficiency, whether it is at the algorithmic level or at the level of chip manufacture. It seems to be a be a natural fit to extend this thinking further into energy consumption as well.

But I have to wonder, how much of a difference can we make? I think that the energy consumption involved in the field of computers - through the whole lifecycle: manufacture, operation & disposal - is relatively low compared to the energy consumed in other areas of the economy (transportation, heating, lighting, manufacture).

Would we be better off spending our time optimizing energy consumption in other industries?

Tax benefit

[slackoon]

The government gives tax benefits for driving hybrid vehicles and I believe they should do it for energy efficient computers as well. "According to the Computer Industry Almanac Web site, at the end of the year 2000, there were 168.84 million computers in use. The projection for the end of 2001 is 182.24 million." So just imagine how many there are now! With that many computers, many of which are never turned off, the energy savings could be enormous.

Re:Tax benefit

[Yartrebo]

Not having a computer is better in terms of energy demand than even having a very efficient computer. Same for the hybrid car. For this reason I find these tax breaks perverse.

Giving tax breaks for efficient items penalizes those who conserve the most by not even having the item or by using less. A business that invests money into writing more efficient software and using less servers should not be penalized vis-a-vis a business that invests the money into more efficient servers.

Performance?

[Speare]

Okay, just stating the wattage is like stating MPG for a car or the energy usage for a fridge. But every year, car performance stays about the same or gets worse, and the fridge ain't getting more full. There doesn't seem to be a single useful energy metric that can drive informed purchasing decisions.

So how do you deal with CPUs that are twice as powerful in the next product cycle? The wattage will be about the same, but the amount you can get done with that chip will be much higher. It's like next year's car suddenly weighs twice as much, or goes twice as fast, or seats two whole families, while getting the same mileage. You can't even consider it in two tiers like "passenger cars vs truck frames" because you have to deal with 2008, 2009, 2010, 2011 performance tiers... they change all the time. How can someone make an informed decision from this?

Re:Performance?

[DMUTPeregrine]

Watts per MFLOP. Or MIPS. Or Watts per Point, where point is an average on some benchmarking system. Just giving watts for a computer is like just giving gallons for a car. You don't know how many miles it can go on those gallons, and so the figure is useless.

*REAL* programmers...

[Anonymous Coward]

*Real* programmers eat microwave popcorn. However, they use the heat from the CPU to cook it instead of using an actual microwave. *REALLY* good programmers can even tell which process is running by the rate of popping.

miles per gallon?

[oyenstikker]

How about "burning library of congresses"

Re:about time.

[Rei]

I agree. For those with home servers like me who leave their computers on all the time, I'd be willing to pay an extra $5 or so per watt saved ($2 for the price extra watt over the course of an assumed usage lifespan of 3 years, $1 for the reduced cooling cost and increase in part lifespan from it being cooler, and $2 extra for the environmental benefit). If I lived in a place with expensive power, like California, that would probably be $7 or so per watt saved.

The marine diesel engine industry

[dj245]

In the marine diesel engine industry, there is a measurement of NOx (nitrous oxides), usually measured in grams per killawatt-hour (g/kW-hr). But not all engines will be used in the same service, so they won't be running at the same load. Some will run at 100% load most of the time they are on (generators, fire pumps maybe) while others will run at about 65 or 75% of full power all of the time- these are your direct-drive propulsion diesels. These different duty cycles have a dramatic effect on the numbers. So what to do?

The International Maritime Organization has created a few different cycles- E2 is Constant Speed Main Propulsion, E3 is Propellor law operated propulsion for example. You pick your cycle, run your engine at a variety of loads, then use weighted averaging on those loads to determine what the emissions would be if the engine ran at E2 all the time. Then you can say that for the E2 cycle, the engine puts out so much NOx.

For computers, someone needs to come up with some different computer cycles. There may be several of them- 50% parallelizable with 25% floating point and 75% integer math, 100% parallelizable with 100% floating point math, etc. Different architectures may take dramatically longer to do floating point or non-parallizable workloads. Only then could you run a bunch of tests and really say that under this load the computer uses this much power to do a certain amount of work in a given amount of time.

This is not new or novel stuff. This is similar to how the EPA tests cars. Some cars do highway miles much better than city miles, so they do both and weight the averages.