Datacenter Power Architecture

Telecom Central Office Datacenters have seen rapid evolution in terms of the architecture of their power supply arrangements; a number of different approaches have been used in relatively quick succession and, indeed, more than one approach may be found within a single installation.

This article considers a number of embedded power supply architectures for Telecom Central Office Datacenters, looked at on a system perspective. Traditionally, power supplies were a simple service function; each major unit within a facility such as a datacenter had its specified requirements in terms of the number and level of its voltage rails, and the current that had to be provisioned in each case. If there were safety and protection requirements – voltage or current limits that, if exceeded, would be regarded as indicating a fault condition – then the response to those conditions would typically be a local trip or shut-down with, at most, a simple flag being set to indicate the existence of a problem.

The hardware that met those requirements changed only slowly over many years; the power supply in the form of the rectangular metal enclosure or “brick” was a familiar system component; self contained, autonomous voltage-in, voltage-out devices. Whether custom-designed or sourced from a standard range, they invariably resided in any available space within – or even outside – the cabinet or rack, but had minimal mechanical and electrical integration with the host system.

Recent years have seen a rapid, and accelerating, pace of change in the design of such central-office systems in terms of their power – both computing or data throughput power, and the electrical power required to deliver it – and in levels of integration, and power density. Power density, in fact, has been one of the fastest-moving parameters, irrespective of whether it is measured at circuit-board, rack, cabinet or total system level, and one of the key challenges in power supply provisioning has been to accommodate and to keep pace with its growth. Alongside, and intimately connected with, higher power, higher density, higher input voltages, and higher load currents sits the overall requirement to achieve higher efficiency. Reducing losses in the power supply sub-system lowers the heat dissipated within the system, with consequent reduction in the burden on cooling arrangements, boosts reliability by reducing component temperatures, assists with improving the environmental credentials of the installation and – far from insignificant at a time of escalating power tariffs – helps to contain rising power bills.

Many of the changes forced on power supply architectures can be traced to the evolution of microprocessors and other semiconductor devices. Successive generations operate at lower voltage levels while consuming higher currents; the processing power they deliver per Watt has grown rapidly but this has not, in general, led to any reduction in power dissipated in a given card or rack. Rather, throughput has risen dramatically while power density has at best grown at a more modest pace. High currents and low voltages force regulation closer and closer to the point of load; otherwise, currents of many tens of Amps, routed over any physically-feasible PCB or wiring structure, suffer voltage drops that are unacceptable as a proportion of a rail that may only be 1V or even less.

The disparity between such low voltage levels at the point of load, and the traditional telecom system-supply rails of (typically) 48V has stimulated an evolution in hierarchical power distribution architectures. For a majority of applications implemented today, the Intermediate Bus Architecture (IBA) has been the preferred power architecture. This power architecture has led to the development of the isolated, semi-regulated DC/DC converter known as the Intermediate Bus Converter (IBC).

Fixed ratio Bus Converters that employ a new power topology known as the Sine Amplitude Converter (SAC) offer dramatic improvements in power density, noise reduction, and efficiency over the existing IBC products. The power systems designer faces the challenge of providing small, cost effective and efficient solutions that keep pace with – or preferably, exceed – the trends in system voltage and current identified above. Traditional power architectures cannot, in the long run, provide the required performance. Vicor’s Factorized Power Architecture (FPA), and the implementation of V•I Chips, provides a revolutionary new and optimal power conversion solution that addresses all of these challenges. The technology behind these power conversion engines used in the IBC and V•I Chips has a direct bearing on the designer's ability to respond to changing power supply demands.

System view
A systematic approach to partitioning shows that several functions need to be accomplished, while each architecture offers different building blocks. System functions are summarized in Table I. These functions present some trade-offs, which need to be considered: for example
• DC step-down ratio vs. efficiency
• Load voltage vs. efficiency
• Battery cells (number and size) vs. backplane voltage range.
• Bus conductor section and current level vs. losses.
The environment in which the typical datacenter operates is also changing, which adds an extra dimension to the matrix a recent trend is towards adoption of High Voltage DC infrastructure distribution instead of standard AC line.

IBA and FPA
The essential features of an Intermediate Bus Architecture (IBA) are that it distributes semiregulated 42-50V voltage (therefore taking advantage of reduced distribution losses); then, by means of an isolated, nonregulated Bus Converter, this voltage is approximately reduced by a factor of 4 (or 5), while the voltage reference is changed from positive to negative (intermediate bus). This intermediate 9-12V bus enables non-isolated Point-Of-Load converters (niPOLs) to provide final step down and regulation function at the same time.

The Factorized Power Architecture (FPA) represents a further evolution onwards from IBA: the high efficiency, fixed ratio, isolated but not regulated converter is used as Point-Of-Load device, providing isolation and current multiplication directly from 48V; upstream, a non isolated device provides voltage regulation by dynamically adjusting the 48V “Factorized Bus” slightly above or slightly below 48V. This avoids the need for the intermediate bus altogether, therefore further increases distribution efficiency.

A state-of-the-art configuration for an IBA for a 48V telecom system is shown in Figure 1. Each stage accomplishes two main functions, as listed in the diagram. An aspect to note about this structure is that the first stage (AC/DC Power Entry Module, PEM) is a high power, highly optimized commodity device, which takes care of power factor correction (often from single or three phase lines) on a universal input voltage line, and provides an isolated output. Similarly, the niPOL comprises a standard, off-the shelf single or multiphase Synchronous Buck converter, which can deal with a relatively wide input range while providing accurate Load regulation. The bus converter is a fixed ratio electronic transformer device that interfaces the 48V backplane with the 9.6V or 12V (nominal) intermediate bus. In terms of total life-cycle cost, all three stages have achieved extremely high performances, with commodity level acquisition costs, and peak efficiencies well in excess of 90%. In order to minimize intermediate bus losses, the IBC is placed as close as possible to the niPOL it supplies.

The dotted lines in Figure 1 also show where the various blocks are located: DC backup and AC/DC PEM are normally stand-alone systems, sized to supply an entire rack through the distribution backplane. IBCs and niPOLs are either discrete or modular devices that are soldered on each unit Mainboard. Further points to note include;
• Having larger, stand-alone AC/DC converters allows lower cost and higher peak efficiency to be achieved.
• IBCs are available as open-frame, through-hole devices with power ratings ranging from few hundred Watts to one kilowatt; as such, they can effectively be placed close to one or a few niPOL regulators.
• niPOL regulators are standard synchronous buck converters that are tailored to each specific load.
The major drawback of this configuration is that blocks at rack level are generally oversized, which implies an added initial cost and higher energy cost because they do not operate at their peak efficiency.

A possible FPA for a typical system is shown in Figure 2. It still consists of three stages of power conversion; while the Adaptive Cell PFM first stage is functionally equivalent to the AC/DC PEM block in the IBA, the load regulation function is now accomplished by a non-isolated converter, the PRM. This converter works with an input line range of 36V to 55V, and outputs a “factorized bus” that can range from 0V up to 55V.

The VTM stage is enabled by the same topology used in the IBC (in effect, an electronic “DC transformer”), and provides both effective current multiplication (by a fixed factor), and isolation. Compared to IBC, FPA offers significant system advantages:
• The factorized bus is normally maintained above 40V, reducing losses and conductor cross-section requirements on the Mainboard.
• The Load voltage that can be regulated has a wider range, given the factorized bus range and the availability of various VTM transformer ratios.
• Just like IBCs, PRMs and VTMs are power components that can be directly soldered on the Mainboard PCB
• The PFM is also a power component that can be mounted on PCB.
In this case also, the dotted lines show where the various components are ideally located. It is clear that FPA offers the possibility of integrating the entire power supply within each unit, potentially avoiding any 48V distribution at rack level.
Given the ever increasing power density of DC backup systems, it is conceivable to integrate that block at Unit level as well, rather than at Rack level. This approach offers a number of advantages;
• Each Switch or Server unit is fully autonomous and does not depend on infrastructure, even just at rack level
• The entire supply system, being more granular, can be tailored more closely to actual power levels of use, therefore maximizing conversion efficiency and minimizing acquisition costs
• Low voltage levels are generated within, and confined to, close proximity of the Load, minimizing distribution requirements and losses.
FPA flexibility is also evident when a high voltage DC supply is available: in this case, by simply substituting the Adaptive Cell PFM block with a High Voltage Bus Converter module (BCM), the same system can be efficiently powered, as shown in Figure 3.

Converter topologies
Three converter topologies contribute to enabling efficient, PCB-mountable power components for both IBC and FPA; these are the Sine Amplitude Converter (SAC); Zero-Voltage-Switching Buck-Boost Converter (ZVS-BB); and Adaptive Cells, Double Clamp Zero Voltage Switching Converter (DC-ZVS).

The Sine Amplitude Converter is a transformer-based series-resonant topology that operates at a fixed frequency, equal to the resonant frequency of the primary side tank circuit, as shown in Figure 4. The entire powertrain (primary and secondary) is locked to the natural resonant frequency of the tank and is switching at primary current zero-crossing points both at turn-on and turn-off, eliminating switching losses in the transistors and reducing the generation of high order noise harmonics. The current in the primary resonant tank is a pure sinusoid rather than a square wave or a partially sinusoidal waveform; because it does not rely on closed loop operations, but on voltage imbalance across its resonant tank, its dynamic performance is effectively broadband, up to 2/3 of the switching frequency.

In response to a load step on the secondary, the Sine Amplitude Converter reacts by increasing the amplitude of the sinusoidal current on the primary resonant tank. The dynamic behaviour associated with this transient is only limited by resistive elements in the conduction path and the time constant associated with filter capacitors' energy depletion, being the primary current driven by voltage imbalance across the tank.
The Zero Voltage Switching Buck-Boost Converter is a conventional buck boost topology with a non-conventional modulation scheme (Figure 5). It is operated in discontinuous mode, with the majority of the power processed with the powertrain inductor directly connected between the input and the output (Phase 2 in Figure 5). Proper switch timing ensures that all the transitions happen at zero voltage, thus eliminating switching losses. The converter can therefore operate at few MHz and easily exceed 97% efficiency.

A unique feature of this topology is the complete symmetry with respect to the input and output line. This device acts as system regulator, taking care at the same time of both input variation, and load regulation. However, in nominal conditions, it is called upon to compensate only for minimal variations, therefore further maximizing efficiency.

The Double Clamp Zero Voltage Switching Converter (DC-ZVS) consists of two cells with the same topology – DC-ZVS – configured dynamically (that is, automatically switched as required) either in series or in parallel, in order to efficiently interface a wide variation of its input voltage. The Double Clamp consist of a standard isolated converter with unconventional modulation and storage elements, which ensure all the switches are turned on either at zero voltage or zero current. This enables switching frequencies in the MHz range with efficiency above 95%. The Adaptive Cell feature allows the converter to operate at the same efficiency over universal input range.

Conclusions
Advances in power conversion technology enable power supply systems to be effectively built and integrated on board as power components. This paradigm shift requires a revision of current system architectures, in order to take fully advantage of power density and efficiency, and reduce total cost of ownership.

www.vicr.com