Jeffrey Skolnick, of the Georgia Institute of Technology, used to just worry about harnessing enough computing power to chase down cancer genes. Today, he’s worried about his electric bill, too.
Skolnick, director of the Institute’s Center for the Study of Systems Biology, in Atlanta, said concerns about server power consumption and heat weighed heavily on researchers who had to design a new 1,000-node computing cluster at Georgia Tech.
“If I’m going to have problems with overheating, and I’m only going to be able to turn on 60 percent of [the nodes], I’m not a happy camper,” said Skolnick, who isn’t the only one fretting about how much electricity his servers gobble up.
Both heat and power have “become a concern in the last 12 months as we have begun to add racks to our new data centers,” said Jevin Jensen, director of IS at Mohawk Industries, in Dalton, Ga. “We are already reworking some [heat and air conditioning units] in one data center to provide better cooling. We are watching this much more closely going forward.”
Chip makers such as Intel and Advanced Micro Devices are also watching. Lower-power chips are expected to be the focal point of the Intel Developer Forum, which kicks off March 7 in San Francisco.
Focusing on energy-efficient chips seems like a good bet. A confluence of events has pushed electric bills to the top of IT managers’ list of worries.
For starters, electric bills rose 10 percent, year over year, to an average of 8.16 cents per kilowatt-hour in November 2005, according to the latest figures from the U.S. Department of Energy’s Energy Information Administration.
Meanwhile, the energy price spike comes alongside an increased number of x86 server rollouts. And it’s unlikely that managers will see any relief soon. A barrel of crude oil, which is coupled with the price of natural gas, continues to hover around $60, forcing electric companies that rely heavily on natural gas to raise rates.
“Electricity costs have gone up so much that [energy consumption is] starting to appear on the plate of the higher-ups,” such as chief financial officers, said Robert Rosen, president of the IBM Share user group in Bethesda, Md., and a member of eWEEK’s Corporate Partner Advisory Board.
“They’re going to start coming back to the IT people and saying, ‘Hey! You’re our biggest power user. You’re costing us x per month. What are you going to do to reduce that?’ IT managers are going to have to start becoming more aware.”
The big question is whether chip makers, which are beginning to tout more energy-efficient processors as a step toward creating lower-power servers, can ride to the rescue.
At IDF, Intel, based in Santa Clara, Calif., will give details about its new server chip platforms and discuss its more power-efficient processor architecture, which is set to arrive later this year. Company officials said the platforms will underscore Intel’s efforts to offer a better balance between performance and power consumption in its chips.
The focus on power consumption marks a change for Intel, which has always sought to boost computer servers’ capabilities, but often by offering faster, more power-hungry chips. Intel now has gone from touting speed to an approach it said could help technology departments hold down electric bills.
Intel officials said the company’s latest server platform, combined with less power-hungry chips, will fend off rival AMD from key accounts such as Google, which, according to Morgan Stanley, is now using AMD chips to save energy. AMD will also highlight its offerings at IDF.
Next Page: Measuring performance per watt.
Along with the unveiling of new platforms will come much discussion of performance per watt, a processor “miles per gallon” measurement designed to illustrate how much work a chip can accomplish for each unit of energy it consumes.
The focus on performance per watt will become increasingly important as corporate data centers continue to add serversdual-core-processor servers, for example, add an extra helping of performancebut struggle to power them and keep them cool.
That means having the ability to size up the total power consumption of a rack full of servers would assist IT managers in choosing new machines, Rosen said.
“What will [power consumption] be if I’m running Intel? What will it be if I’m running AMD? What will it be if I’m running [IBM Power5]? That would really be interesting as we move down the line where we really have to worry more about power,” Rosen said.
“If you can say, ‘For my application, any of these will work,’ then with all other things being equal, why would I not want to go with the lowest-power system? Because that pays me back over the life of the system.”
Winning over users such as Jensen and Rosen is becoming a top priority for the chip makers, which intend to roll out major server chip platforms this year that limit power consumption, while still incorporating higher-performing multicore processors, along with enhancements such as built-in virtualization and faster memory.
Businesses will be able to use the chips to deploy higher-performance servers that can be packed more tightly into space-constrained data centers without necessarily prompting upgrades in electrical service or cooling, chip makers said.
Intel’s power plan
Bent on making technology managers stick with its chips, Intel will detail its new server platform, code-named Bensley, at IDF. The company also is expected to announce that it has begun shipping its dual-core “Dempsey” Xeon DP chips.
Servers based on the platform, which is designed to bump performance by doubling up on buses that carry data to processors and memory, as well as by incorporating FB-DIMMs (fully buffered dual in-line memory modules), should arrive in May.
Bensley also includes Intel’s Virtualization Technology and I/O Acceleration Technology, which make it easier to divide a machine to run different software and to increase network throughput, respectively. The add-ons are designed to increase server utilization, offering another way to cut power by allowing a company to use fewer servers and to run them harder.
Bensley will work with only the 65-nanometer Dempsey chip at first. Intel officials said the chip offers a better mix of performance per watt, by bumping performance and coming in a lower-power variant. FB-DIMMs, however, will use more electricity.
However, when combined with “Woodcrest,” a more energy-efficient dual-core chip due in the second half of 2006, Bensley will offer much greater performance per watt, allowing businesses to save energy by adopting the platform, said Kirk Skaugen, general manager of Intel’s Server Platforms Group, in Hillsboro, Ore.
“Woodcrest in the Bensley platform … will have a 3X improvement in performance per watt versus the volume platform we were shipping in 2005,” Skaugen said, referring to Intel’s single-core “Irwindale” Xeon chip.
Next Page: AMD fires back.
AMD officials were quick to refute claims such as Skaugen’s. AMD chips, they said, offer advantages to businesses.
“What they want to do is take their existing data centers and pack more compute density into those data centers,” said Randy Allen, corporate vice president for AMD’s server and workstation division, in Austin, Texas.
“The reason Opteron has been successful is that, within the same square footage of a data center, within the same power infrastructure, we’ve been able to deliver vastly more performance.”
AMD will demonstrate a four-processor machine based on its newest dual-core Opterons at IDF.
The new chips, based on a revised design AMD calls Rev F, will touch up AMD’s Opteron server platform by offering features such as built-in virtualization and an on-board memory controller that addresses DDR2 (double-data-rate 2) memory.
Access to DDR2 memory modules will afford servers incremental performance gains if companies choose the fastest DDR2 800 modules.
AMD officials said the way the company has constructed its chip platform chips link directly to each other and use an on-board controller to tap into memorygives it a major power advantage over Intel. AMD officials said the core chips inside servers using its Opteron Rev F chips will use about 100 watts less than those in Intel Bensley/Dempsey machines.
An Intel spokesperson said the difference between the companies’ chips would be much less than the 100 watts AMD claims, especially following Woodcrest’s arrival.
There are no universally accepted tests, analogous to a fuel mileage rating, that allow a company to compare one machine against another. “The industry is looking for more objective ways to measure power consumption,” said Alex Yost, director for xSeries servers in IBM’s Systems & Technology Group, in Somers, N.Y. Intel, for one, would support such a measurement, according to Skaugen. However, it would have to be developed by third parties, he said.
“I think that the answer is going to be harder to come by,” Yost said, instead prescribing what he called good diet and exercise for servers. “It’s about making sure, as you deploy your server platform, you maximize utilization,” he said. That includes taking advantage of new technology such as virtualization.
“All I have to go on now is watts and Btus,” Jensen said. “We do monitor our server CPU temperature closely. But it is too late at that point. You have already purchased and installed it. It would be great to know in advance.”
Additional reporting by Senior Editor Jeffrey Burt
Check out eWEEK.com’s for the latest news, views and analysis on servers, switches and networking protocols for the enterprise and small businesses.