There's a new opportunity to design added-value power supplies and chips that help data centers conserve electrical power. The problems that IT managers face today are nothing less than monumental. Consider a historical analogy.
Fast forward two decades: Seventy miles upriver from Bonneville, the Dalles Dam, with a present capacity of 1.78 GW, was completed in 1957. Last summer, Google announced it would build its next data center in the Dalles, Ore., to be close to all that electricity.
Also, Microsoft and Yahoo are building their own data centers further up the Columbia at Wenatchee and Quincy, Wash. (near the Rocky Reach and Wells dams—which represent 1.4- and 0.840-GW capacity, respectively). That's right. Today's data centers use electricity as if they were aluminum smelters.
Andrew Nebbs of EYP Mission Critical Facilities described a standard half cabinet containing six PowerEdge 1855 7U high chassis, each containing 10 Dell Dual Xeon Server Blades, at Darnell's Digital Power Forum last September. Each chassis in a cabinet has a pair of power supplies rated at 2.1 maximum kW. That's 4.2 kW per chassis and 25.2 kW per cabinet.
That's not sniping at Dell, though. Nebbs came up with similar numbers for Sun. And looking ahead, other presenters at the forum saw future-generation half cabinets consuming up to 40 kW. Of course, that's only the electronics. Now consider the cooling load.
So it's no wonder Google, Microsoft, and Yahoo want to be next to their own dams. Okay, that last part is exaggeration. Nebbs and other speakers said they want to be close to the dams to minimize power transmission losses and to reduce their exposure to grid blackouts. But those concerns reflect the enormous amounts of power they will be consuming.
Chip companies in the power market have picked up on the challenge. In one talk, Dave Freeman of Texas Instruments showed how closing the power control loop in the digital domain could allow adaptive optimization of the feedback; "adaptive" in the sense of adjustment for transient, component, and environmental changes.
Freeman's talk focused on adaptation within the power-conversion electronics down at the power-supply or voltage regulation level. But adaptation works in other ways as well.
For example, other presentations at the forum dealt with "virtualization" of server tasks. To date, server operating systems have been designed to distribute the processing load across all available servers. Virtualization would make it possible to shift processes dynamically so some server blades could be shut down at off-peak times. This is critical to conservation because the load presented by each active blade tends to be constant, regardless of how hard it's working.
One way to facilitate virtualization is via the two-way communication made possible by digital power-supply control and the PMBus digital communications architecture for power chips and products. In its simplest form, that communication might be a matter of each of the many supplies in the cabinet reporting their current load conditions to the operating system.
More ambitiously, Zilker Labs' Jim MacDonald proposed thermal management within each cabinet by taking advantage of the local temperature-reporting capabilities provided by the half-dozen or so digital dc-dc converters on each server blade. With temperature data from each converter in the cabinet reported via PMBus, a system controller in each cabinet would have a nearly real-time picture of heat distribution within the cabinet, enabling it to shift loads and control cooling fans to maximize efficiency.