Sidebar: The big drain: where the most power is wasted

Computerworld - Increasing power densities of networked storage, communications equipment and servers all contribute to the power and cooling problem. However, servers have become the biggest issue because server farms have grown so large relative to everything else in the data center. Within the server, increasing power budgets for processors are one of the biggest problems, says Nathan Brookwood, principal at Insight64. "When the first Pentium chip came out, everyone was horrified because it used 16 watts," substantially more than the 486, he says. Today, the worst offenders are at 150 watts, and 70 to 90 watts is common. Processor inefficiencies reached critical mass when chips moved to 90nm technology. "The leakable current for transistors became a meaningful part of overall power consumption," he says. As heat loads were going up, increasing voltage and frequency was no longer an option.

"Power was going up at an alarming rate," says Randy Allen, corporate vice president at chip maker Advanced Micro Devices Inc.. New dual-core chips from Intel Corp. and AMD may give IT a bit of breathing room. Intel says its Woodcrest generation of processors, due later this year, will cut thermal design power by 40% over current single-core designs, while AMD claims its dual-core chips offer 60% faster performance within the same "power envelope" as its single-core chips. Both also offer capabilities to step down power levels when utilization levels drop. For example, AMD claims that a 20% drop in performance level will cut power levels on its newest chips by 50%.

While dual-core and future multicore designs may give data center managers a bit of breathing room, they won't stop the power-density crisis. While performance per watt is going up, total wattage per chip is staying at the same or slightly higher levels and will continue to climb. Allen likens the battle to trying to climb a falling ladder. "The fundamental issue here is that computational requirements are increasing at a faster rate" than efficiency gains are rising, he says.

Beyond the processor, there are areas of inefficiency within the server itself. For example, vendors commonly put inefficient power supplies in high-volume x86-class server because they don't see a competitive advantage in putting in more-efficient components that add a few dollars to their server costs. Jon Koomey, consulting professor at Stanford University who has studied data center power issues for the industry, calls this a "perverse incentive that pervades the design and operation of data centers." Commonly used power supplies have a typical efficiency of 65% and are a huge generator of waste heat. Units with efficiencies of 90% or better will pay for themselves in reduced operating costs over the life of the equipment. There's a second benefit as well, says Koomey. Data center managers who are purchasing thousands of servers could also design new data centers with a lower capital investment in cooling capacity by purchasing more efficient servers. "If you are able to reduce power use in the server, that allows you to reduce the capital cost of the equipment and you can save costs upfront," says Koomey. Power supplies are a simple way to gain efficiency, he adds, because they can be inserted into existing system designs without modification.