UPS

Efficient architectures move sources closer to loads

Editor's Note: This feature is part of a special section on new power paradigms. For more, see the related articles at bottom or go to Power Management DesignLine.

Cars and heavy industry get the headlines, but the pristine IT data center hides an enormous consumer of energy that is rapidly increasing its carbon footprint. Consuming enough energy to power as much as 25,000 to 35,000 homes, these centers are alarming accounting departments and raising the hackles of green activists worldwide.

At the heart of the problem is inefficient power delivery, specifically power conversion, memory leakage, cooling and distribution losses. In recognition of this, the world's top engineers have scrutinized everything from power distribution to server and point-of-load power supplies to minimize power waste. Their work has led to architectural breakthroughs that look to fundamentally change the face of power delivery for decades to come.

Commenting on the problem of efficiency, Stephen Oliver, vice president of marketing and sales at Vicor (Andover, MA), said "For every watt of power used usefully you have to put 2.3W into the building just to get there." Oliver was essentially referring to a metric called power usage effectiveness, or PUE. PUE was defined by an organization of IT professionals called The Green Grid that was formed to promote energy efficiency in data centers. PUE is the ratio of total power used by a data center divided by the IT equipment power.

Click on image to enlarge. The power chain in a typical U.S. data center has two primary obstacles to efficiency: the number of power conversions in the chain and the distribution loss.

The aggregate inefficiency comes as a result of all of the power conversions in a data center and the efficiency of components such as processors, memory and disk drives. For example, a server power supply might offer 90 percent efficiency. That 10 percent of energy that is not used effectively adds to the PUE. Moreover that 10 percent of wasted energy is dissipated as heat. The power required to cool the equipment is also captured in PUE.

Dallas Thornton, division director at the San Diego Supercomputer Center (SDSC) located at the University of California San Diego (UCSD), believes data centers can hit a PUE goal in the 1.3 to 1.4 range. Thornton is currently working on an addition to the SDSC. He states, "Typical data centers today are in the 1.8 range. Inefficient data centers can be over 2."

To improve efficiency, the industry needs more-efficient power supplies and improved power-systems management. The industry must also consider more efficient power-distribution systems. And certainly the efficiency of the cooling scheme must be optimized. With the exception of cooling systems, all of the other efficiency initiatives rely on engineers adapting and optimizing power supply and server designs.

The figure depicts a simplified power chain in a typical U.S. data center. An ac input feeds an uninterrupted power supply (UPS) typically at 480-V phase-to-phase, 277-V phase-to-neutral levels. The UPS outputs the same voltage level to a PDU (power distribution unit) that includes circuit breakers and a transformer that steps the voltage down to 208 V. The server PSU (power supply unit) converts the 208 V input to 12 Vdc.

The power system shown has two primary obstacles to efficiency. First, there are a number of power conversions in the chain. Bill Tschudi, program manager at Lawrence Berkeley National Laboratory (LBNL), states "Every time you do a conversion from ac/dc or vice versa, there is a loss. That loss becomes heat and that heat has to be taken out." In fact, the typical double-conversion UPS design first converts ac to dc to charge the battery, and then back to ac to feed the PDU. The PDU transformer adds to the loss as does the PSU that does another ac/dc conversion.

The second obstacle is distribution loss. The PDU outputs a relatively low 208-V level. And while the PDU looks simple in the diagram, in reality it may feed many racks of servers. A relatively low voltage means the system needs relatively high current. High current is bad because distribution loss is driven by the square of the current. To maximize efficiency, Vicor's Oliver states "Keep the voltage as high as you can as far as you can."

Tschudi from LBNL has done research sponsored by the state and federal governments and utility companies about moving to a 380-Vdc distribution scheme to both attack distribution loss and minimize the power conversions. In such a system, the UPS would supply 380 Vdc directly to what would become a dc/dc server PSU. A PDU would still include circuit breakers but perform no power conversion. Tschudi states, "If you eliminate both the conversion at the UPS system and the conversion at the first stage of the power supply, you've gotten to the 9 pecent range or so." He is referring first to the dc/ac conversion at the output of the UPS, and the ac rectification stage in the PSU. He has run tests that show a 9 percent energy savings over what he claims are the best ac systems.

The 380-V level was chosen because it's relatively high compared with 208-V dc, and it would be a good match for server PSU designs. Most 208-V input ac/dc PSUs generate a voltage in the 350- to 380-Vdc range after the initial rectification stage. So PSU designers could presumably just eliminate that rectification stage from existing designs.

Not everyone agrees with Tschudi's research. The UPS vendors in particular claim that Tschudi did not use the latest most-efficient products. Gary Anderson, marketing manager at Emerson Network Power (St. Louis, MO), states "3 to 5 percent is pretty much the max" in terms of efficiency gain from a 380-Vdc scheme. He also defends the power transformation in the PDU as being important from a reliability point of view because it adds an isolation function. Anderson states, "Customers are looking for reliability as much as they are efficiency."

There are also logistic issues with a dc system. Anderson points out that there aren't regulatory codes in place for the dc scheme, nor are their components such as circuit breakers and connectors. Tschudi counters that applications such as mass transit and elevators use dc power and that suitable circuit breakers exist. Tschudi has also cooperated with a number of organizations worldwide that are interested in the dc scheme and the informal group has agreed on a connector that could be used globally. Presumably, Anderson Power Product will introduce a version of the connector shortly.

Deciding whether a 380-Vdc scheme will deliver the promised efficiency comes down to whom you believe. The Green Grid did a study that compared eight different ac and dc schemes and found them all within a few percentage points in efficiency. APC (American Power Conversion), another UPS vendor, has published several whitepapers on the subject and puts the dc advantage in the 1- to 2-percent range. Tschudi believes there will be real deployments to benchmark soon. Nippon Telegraph and Telephone (NTT) is working on a dc data center in Japan. And LBNL will work with UCSD on side-by-side tests using Sun equipment later this year.

APC, meanwhile, is proposing a slight alternative to the ac system. Victor Avelar, senior research analyst, states, "Do what you can to get to 240-ac line-to-neutral." Avelar claims that even raising the distribution voltage from 208 Vac to 240 Vac would offer some advantage in minimizing distribution loss. Moreover he points out that 240 V is the highest ac voltage to which existing technologies could easily adapt. He claims that server PSUs will accept 240V maximum as an input, and that existing electrical connectors are only rated to 240 V. APC advocates stepping down to 240 V via an autotransformer that is more efficient than the isolation transformer in most PDUs.

Were it not for the 240 V limit, data centers could presumably distribute the 277 V line-to-neutral power directly from the UPS. In fact IT managers would like to buy 277V-capable gear. According to UCSD's Thornton, he will buy some specialized computers from vendors such as Sun and IBM that run directly from 277-V service. But mainstream systems with commodity blade servers don't offer that option.

Designers working on server power system architectures clearly need to consider the possibility that high-voltage inputs may be coming their way. At the same time, the designers are faced with building intelligent power systems that can communicate with the power-management software running on the server. Tim Phillips, vice president of Enterprise Power at International Rectifier (IR) (El Segundo, CA) states, "It's not just making your power supply more efficient, but allowing your power supply to be controlled that makes the digital electronics more efficient."

Phillips is implying two-way communications. The fact is that any part of a data center--the UPS, the PSU, a server, a processor core--rarely runs at capacity. Each subsystem and component is typically architected to handle worst-case loads with some redundancy for continued operation when something fails. Typically, everything operates at 35 percent or so of capacity. IR supports point-of-load supplies that feature phase shedding to boost efficiency at low load levels. Such a converter uses parallel phases to supply peak power yet can disable a phase for typical levels.

Phillips also believes that the power-management system must manage the processors so that no core goes into thermal shutdown. Even when processor demand is high, it's better for management software to communicate with the power system and perhaps reduce processor voltage 10 or 15 percent, thereby decreasing performance as well, rather than having the processor overheat and temporarily shutdown.

As for handling higher voltages, the power supply and power component must remain agnostic to any scheme such as 380 Vdc, while preparing to support such a technology. Vicor has proposed an architecture that uses what it calls a bus converter module (BCM) to convert 380 Vdc to 48 Vdc. The BCM is in the industry standard one-sixteenth brick form factor. The company has a 1-square-inch voltage transformer module that can convert from 48 V to sub-1 V processor input levels right at the load.

Analog Device also has a 380-Vdc-input reference design using its ADP1043 digital PWM controller. The reference design outputs 12 V in the type of configuration proposed by LBNL's Tschudi.

Perhaps the biggest change coming for power-supply design is who takes responsibility for that design. Looking back, system houses have depended on dedicated power-supply vendors for so-called silver-box products such as server PSUs and for dc/dc converters. These days many dc/dc converters are essentially chips--at least in terms of package--and more of the power system is being integrated onto the server board. Vicor's Oliver points to an IBM design that uses 17 of the companies BCMs delivering 5.1 kW of power directly on a server board. Phillips' points to a server design that integrates 86 IR components on the PCB. Without question architecture is taking higher power closer to the load.

---

Maury Wright is editor of TechOnline's Power Management DesignLine.

Related Links:

MOSFET

Linear Technology

Power MOSFETs

Vishay Intertechnology

STMicroelectronics

Batteries

Power MOSFET

ADC

Solar

Diodes

Photovoltaic

Analog Devices

Battery

Smart Grid

Maxim Integrated Products

Power Supplies

Cypress Semiconductor

National Semiconductor

MOSFETs

Fairchild Semiconductor

International Rectifier

DC/DC Converter

IMS Research

Texas Instruments

Microcontroller

Power Supply

Power

Power Management

Energy Harvesting

Analog