Power demands of the latest microprocessors from AMD, IBM, Intel, and other vendors continue to soar. Indeed, CPUs currently under development for next-generation servers require 150 A or more, while desktop processors require 120 A. Along with these stringent requirements come additional constraints such as cost, board area (power density), and thermal environment, which further inflame the design challenge.
To explore the various options for voltage-regulator design, this article will review several multiphase buck voltage-regulator-down (VRD) solutions that target the latest CPUs in server applications. A thermal design current (TDC) of 130 A, with a peak current of 150 A, will be used as a design target. The evaluation will be based on the use of an eight-layer pc board or PCB, which is commonly implemented in server motherboards. The power and ground planes use 2-oz copper (middle layers), while the rest of the layers are 1-oz copper.
The thermal challenge
When delivering continuous currents of 130 A at 1.35 VOUT, an 85% efficient VRD design would need to dissipate about 31 W to the ambient—a requirement that makes the thermal design a formidable task. In rack-mounted server environments, ambient temperatures inside the chassis are as high as 45°C, further raising the performance bar. In such a system, saving every watt of power loss is important because the heat-saturated PCB has a high junction-to-ambient thermal resistance. PCB temperatures generally must be kept below 105°C to maintain reliability. In most cases, this is the factor that restricts the amount of voltage-regulator power loss, rather than the component temperatures that generally can be much higher.
The thermal design problem can be solved either by increasing the voltage regulator's efficiency, by finding a more effective way to remove heat from the PCB, or by doing both simultaneously. For a given set of controllers, drivers, and passives, efficiency improvements can be achieved quite simply by increasing the number of MOSFETs paralleled or by upping the number of phases in a design.
Having more paralleled MOSFETs helps reduce MOSFET conduction losses, but it doesn't help reduce the stress on other components, such as the inductors. Paralleling more MOSFETs per phase also increases layout complexity and switching losses, limiting the design's switching frequency. Furthermore, paralleling MOSFETs increases the area of the switch node (the connection point of the MOSFETs and inductor), which causes an increase in radiated noise. Because high-speed digital lines usually must be run under or near the switch nodes, this can cause interference and prevent proper motherboard operation. For these reasons, the VRD designs considered in this article will avoid the use of paralleled MOSFETs.
Boosting the number of phases helps reduce stress in every component. Moreover, it's possible to use smaller and less expensive components, reduce ripple current in I/O filters, and improve transient response. However, with thermal design currents typically exceeding 100 A, designs are already utilizing all four phases provided by commonly used multiphase control ICs possessing four MOSFETs per phase. Recently, multiphase controller manufacturers have introduced controllers capable of four-, five-, and six-phase operation. But these are based on existing architectures with a fixed number of phases that require point-to-point wiring between the phase circuitry and the controller.
The problem could be solved with a multiphase architecture that lets five, six, or even more phases be implemented while minimizing layout complexity. The IR3081 XPhase control IC and IR3086 phase IC permit such an architecture by including all of the "one-per-converter" circuitry in the control IC and the "one-per-phase" circuitry in the phase IC. The control IC communicates with phase ICs via a novel five-wire analog bus that consists of bias voltage, phase timing, average current, error-amplifier output, and voltage-identification (VID) voltage. By eliminating the need for point-to-point wiring between the controller and the phase ICs, the five-wire bus shortens the interconnects to cut parasitic inductance and noise.
This architecture can support any number of phases, and it facilitates rapid design tradeoffs. Thus, designers can upgrade their designs if a higher-current CPU is introduced or depopulate a phase or two if the current requirements drop. As will be discussed later, this can be done without changing the fundamental design.
High-density, low-profile, server VRD design
Many high-density server designs with two or more processors and closely spaced VRDs require efficiencies in excess of 85% to reduce the generated heat. Height restrictions or the need for unrestricted airflow to the processors might prevent the use of a heatsink. To solve this problem, a seven-phase VRD demo board exhibiting extremely high density and efficiency was created.
The architecture used for this demo board allows for flexibility in designing multiphase, interleaved buck converters with 1 to X phases, while providing the ability to facilitate rapid design tradeoffs. It consists of an IR3081 control IC and a scalable array of phase blocks, each using an IR3086 single-phase IC (Fig. 1). The control IC communicates with the phase ICs through the five-wire analog bus.
The control IC contains the entire one-per-converter circuitry, which includes VID, pulse-width modulation (PWM) ramp oscillator, error amplifier, bias voltage, fault detection, and other necessary functions to meet the load line, as well as the transient response requirements of present and future microprocessors. Each phase IC comprises all one-per-phase circuitry that includes gate drivers, PWM comparator and latch, overvoltage protection, and current sensing and sharing.
Figure 1 also shows the demo board developed using the XPhase architecture and the IRF6617/IRF6691 DirectFET MOSFET pair. The IRF6617 features low gate charge (11 nC typical) for fast switching, while the IRF6691 provides an extremely low RDS(ON) (1.2 mΩ typical). To minimize switching losses, size, and cost, only one IRF6617/6691 pair was used per phase. The control IC is located to the right bottom corner of the board, far away from the heat and noise of the power train. One DirectFET pair per phase and small (10- by 7-mm) Pulse PAO511 220-nH inductors help shrink the power train to 2.5 by 0.95 in. This is considerably smaller than what's possible in four-phase designs that use four MOSFETs per phase and much larger inductors.
A constant 400-kHz switching frequency per phase achieves an equivalent ripple frequency of 2.8 MHz. This frequency was chosen as a tradeoff between overall converter size, transient performance, and efficiency (power loss). The Pulse inductors have a small footprint and low power loss due to their low dc resistance of 0.36 mΩ typical. But they also have a low 32.5-A saturation current that prevents using a lower switching frequency.
A lower switching frequency could yield higher efficiency. But it would require using a higher-value output inductor with a larger footprint, and it could possibly expand the size and space needed for input and output decoupling capacitors. A higher switching frequency could make for a more compact design, but it would increase switching losses and lower the efficiency.
Using seven phases with small inductor values improves the load transient performance, eliminates the need for bulky electrolytic capacitors, and greatly reduces the required PCB area. This design meets the transient response requirements for a load step of 100 A using both ceramic (60- by 10-µF 1206 MLCC) and low-profile POSCAPs (8 by 2R5TPE470M9 470-µF POSCAP with 9-mΩ equivalent series resistance).
A simple linear regulator with an npn pass transistor is used to supply a gate-drive bias of about 7.6 V from the 12-V input rail. Selecting a gate-drive voltage between the traditional 5- and 12-V supplies lets the designer optimize the gate drive for the lowest power losses (Fig. 2).
The seven-phase VRD was evaluated in a wind tunnel at 45°C and experiencing varying amounts of airflow, with and without the heatsink (Fig. 3). Power loss for the design was approximately 24 W, and efficiency is about 87% at 130 A. With the heatsink, the highest temperature on the VRD was about 83°C (Fig. 4). Alternatively, less airflow is required to cool the seven-phase design or a higher ambient temperature is possible. Without the heatsink, more airflow is required to maintain safe temperatures (Fig. 5). But the VRD height is reduced to a maximum of just 5 mm, which can be an important factor in space-constrained applications such as blade server designs, where height restrictions limit the use of heatsinks.
It's possible to reduce the number of phases in applications that have space for additional bulk capacitors. However, power losses increase and a heatsink becomes necessary to remove heat from the board. Due to the flexibility of the multiphase architecture, the seven-phase design was easily converted to a five-phase design by simply depopulating components from two phases.
A finned heatsink, measuring 3.25 in. long, 0.5 in. wide, and 0.5 in. high, was attached to the PCB by four screws extending up from the backside of the board. An adhesive thermal tape (gap pad) was used between the heatsink and DirectFET devices as a thermal interface and an electrical insulator. Other attachment methods and alternate thermal interface materials can easily be accommodated, depending upon the designer's preference.
In this design, the gap pad material used was Berquist A3000 with a thermal conductivity of 2 W/mK. Larger inductors (Pulse PAO515, 225 µH, dc resistance = 0.63 mΩ) with a higher saturation current of 55 A had to be employed because of the higher per-phase current. Switching frequency was 400 kHz.
The five-phase design was evaluated in a wind tunnel. Performance data from the testing is presented in Figure 6. Efficiency of the five-phase design is a couple of points lower at 130 A, and it dissipates 27 W of power loss. This 12.5% higher power loss compared to the seven-phase design is all it takes to drive board temperatures up another 10 degrees. Because most server motherboards have at least two CPUs per board, this greater power loss necessitates a heatsink to help cool the PCB.
Besides the reduction in thermal performance, transient performance is also affected. This design would need 12 to 16 560-µF electrolytic capacitors, which would consume board space and increase costs.