For the PDF version of this article, click here.
Throughout the telecom slump of the last few years, the technology laying the foundation for the information superhighway has continued to advance at an incredible pace. Core voltages of field-programmable gate arrays (FPGAs), network processors and microprocessors are at the 1-V level and going down.
However, the lower voltages are creating problems on the power management side, because most of the voltage busses available to create these low voltages are between 2.5 V and 12 V. The intermediate bus architecture (IBA) scheme and the distributed power architecture (DPA) scheme rely on a medium-voltage bus that needs to be stepped down locally through point-of-load conversion, which is typically accomplished by switching regulators.
These architectures put a heavy demand on switching regulators to create the 1-V rails with low output-voltage ripple in order to maintain the integrity of the core voltage of the processors and avoid data errors in high-reliability networks. However, there are legacy solutions such as synchronous buck and multiphase buck regulators, as well as new architectures using very high-speed post-regulation to get the enhanced performance needed from modern digital processing cores.
New processor core voltages are beign driven lower and lower in the ever-increasing quest to improve performance while maintaining or decreasing power consumption. The 1-V level has been reached on mainstream processors, and the industry is seeing that voltage drop even lower as new ICs developed on advanced process technologies such as 90 nm and 65 nm are released. The problem that arises with these low voltage requirements is that the tolerance on the supply voltages is remaining a fixed percentage of the supply voltage, instead of being an absolute value in volts. This means that as the core voltage goes down, the amount of tolerance that the processor can handle becomes smaller and smaller. For example, some processors have a total accuracy requirement of 610% on the main core voltage. This includes all error terms: dc accuracy, ac accuracy and losses in the printed-circuit board (PCB).
DC accuracy simply accounts for the variation of the power supply relative to all things steady state. The regulation component will have some accuracy term that will vary from part to part, and it is normally specified by the manufacturer and tightly controlled. Also included in this dc term are offsets on the output voltage caused by changes in the constant output current or line voltage. Load regulation and line regulation measure the variation on the output voltage of the regulator relative to a change in output current or the input voltage, respectively. The combination of the voltage accuracy over temperature plus line and load regulation gives the total dc accuracy of the regulator.
AC regulation is critical to the overall tolerance of the processor due to the current draw of most processors being very volatile. A changing current on the output causes a changing voltage on the output voltage. The voltage response of the output depends on the output impedance of the power supply (Fig. 1). If the power supply has very high impedance at frequency, then the voltage response to a current change at that frequency will be very poor. A change in current over a fixed period of time multiplied by the output impedance of the power supply at the corresponding frequency gives the instantaneous voltage droop, or ac voltage change seen on the output of the power supply. Lowering the output impedance reduces the voltage droop. The major factors to output impedance are: the capacitors on the output of the power supply and the impedance of the power supply itself. The impedance of the power supply is a direct indication of the gain versus frequency of the loop and, therefore, the gain bandwidth of the loop.
This discussion regarding the ac response is independent of the type of power supply being used, whether linear or switch mode. An additional contributing factor that is radically different between the two types of power supplies is the output ripple voltage. By itself, a linear regulator does not generate any significant ripple. But, by its very nature, a switch-mode converter creates a significant amount of ripple voltage. A synchronous single-phase dc-dc converter that is fully loaded can have anywhere from 10 mV to 100 mV of voltage ripple on the output. Relative to a 1-V output voltage, that is 1% to 10% of the regulator output voltage. Additionally, switch-mode converters in general have slower response times to load changes compared to linear regulators. This means that the ac voltage regulation also will be worse compared to a linear regulator. Therefore, the combined voltage ripple and ac response of a switch-mode regulator can violate the tolerance requirements of the processor without even considering the dc contribution or the PCB contribution.
At first glance, the contribution to the voltage error caused by the PCB doesn't appear to be significant. However, at currents greater than 1 A, that error term increases significantly as a percentage of a sub-1-V output when not properly considered. When the power supply is located far from the load, there can be a drop in the PCB due to the resistance of the traces in the PCB multiplied by the current drawn by the load. The PCB trace factor is easily managed through true Kelvin sensing at the load, although it won't eliminate the term completely. The closeness of the Kelvin measurement to the power supply pin — or even the contribution of bond-wire losses from the pins of the package of the processor to the bond pads of the silicon itself — can affect the value of the voltage that actually reaches the processor. This can affect the overall performance. The PCB factor is not the primary contributor to the overall tolerance of the output of the power supply, but it does put additional strain on the dc and ac accuracy of the power supply in order to meet the overall voltage tolerance requirements of the processor.
AC and DC Components
Next, let's turn our attention to the ac and dc components because these are the major contributors to the overall tolerance equation. And if these components can be handled effectively by the power supply designer, the PCB losses can be accounted for with ease.
The ac losses are our major concern when considering the design of a power supply for processor applications. Direct linear regulation to the core voltage from an intermediate voltage is not possible, because the current levels and the input-to-output voltage differential typically make this a highly inefficient and therefore undesirable heat-generating approach to power management.
Most telecom and networking applications use a fairly high power-consumption processor or a processor that will generate plenty of heat on its own; therefore, the power supply should be as efficient as possible to reduce the amount of heat generation. Switch-mode conversion is an absolute necessity. The most frequently used topology is the synchronous buck converter. This converter samples the input voltage by switching the input voltage to the output at a fixed duty cycle that depends upon the input and output voltage and the loss terms. This switched input voltage is filtered through a low pass filter to give a dc output voltage with an ac ripple term.
The ac ripple voltage is a function of the switching frequency and the characteristics of the filter. The ripple voltage can be calculated as a function of the amount of capacitance, along with the ESR rating of that capacitor, the ripple current of the converter and the switching frequency. The higher the switching frequency, the easier it is to reduce the switching ripple with small-sized filter components.
Higher switching frequencies come with a downside, however, as they cause a higher amount of power loss in the main switching elements due to the charge necessary to switch the devices on and off. This charge multiplied by the switching frequency gives a power loss term that is directly proportional to frequency. In a synchronous converter, there are two switches with a similar amount of gate charge each, meaning that the gate-charge losses are doubled. Therefore, switching at a very high frequency in order to reduce output ripple is not an efficient solution.
Striving for Efficiency
The typical method for achieving the efficiency necessary and reducing the output ripple is to use multiple converters in parallel, running each converter's switching frequency out of phase with each other. The theory is that if multiple converters are used and running at the same frequency, but are simply out of phase with each other, the amount of average ripple on the output of the converter can be reduced (Fig. 2).
The obvious downside of this approach is the cost of adding a second converter to the main device to create a low-ripple output voltage. A new approach that solves the problem of voltage ripple while maintaining a low-cost approach is to take advantage of the benefits of a linear regulator while offering the efficiency of a switching regulator. A standard single-phase dc-dc converter can be used to create a low voltage such as 1.2 V. This is created very efficiently with a reasonable amount of ripple, say 50 mV to 100 mV. Then, a second stage of power conversion is done with a linear regulator directly at the load. The linear regulator must possess the following characteristics in order for this approach to be effective:
Good power supply rejection ratio at the switching frequency of the dc-dc converter.
Large signal bandwidth so that it can respond quickly to load transients.
Low dropout voltage so that the afore listed characteristics are present at low input-to-output voltage conditions.
In order to provide a low-voltage ripple supply to the core of a processor when post-regulating from a dc-dc converter, the post-regulator must be able to attenuate the ripple generated by the dc-dc converter. Therefore, the power supply rejection ratio (PSRR) of the post-regulator must be good even at typical switching frequencies of 300 kHz, 500 kHz and 1 MHz. The most typical switching frequency used is about 500 kHz. This frequency offers a good balance between efficiency and external component size. PSRR is a measure of the amount of noise from the input that is rejected before reaching the output. Therefore, if the post-regulator has high PSRR, then the switching ripple at 500 kHz can be attenuated significantly from the input of the post-regulator to the output of the post-regulator.
The load transient response of a linear regulator can be made to be much higher than that of a dc-dc converter because it is a linear system. With the proper loop design, the output of the linear regulator can have small signal gain-bandwidth products greater than 5 MHz and large signal bandwidth response greater than 1 MHz. A typical dc-dc converter is limited by the switching frequency. In order to remain stable under all conditions, the dc-dc converter typically needs to roll off its gain and operate with a gain-bandwidth product of one-quarter of the switching frequency. A dc-dc converter with a gain-bandwidth product that approaches the switching frequency may try to correct for the switching ripple on the output, causing an unstable system. Therefore, the switching regulator approach is inherently unable to reach the same response time as a linear post-regulator.
This response time is important because changes in the load current in telecom and networking equipment can be very fast. These fast changes in current cause the output of the power supply to change, so the output power supply needs to react quickly to pull the output back into regulation before causing any kind of reset pulse on the processor side. Whether linear or switch mode, the regulator itself cannot do all the work; it relies on the bulk capacitance as well to hold up the output voltage until the regulator begins to respond. A high-bandwidth regulator needs the output held up less time when compared to a lower-bandwidth regulator. This will reduce the amount of capacitance necessary to maintain the same voltage tolerance.
The Ideal Solution
An ideal solution for powering low-voltage processors in a high-reliability environment is a dc-dc converter post-regulated with a high-bandwidth linear regulator. This gives the best combination of efficiency, low noise and fast response to load changes as well as being able to locate the post-regulator directly at the load that is being powered. Micrel's MIC5190 is an example of a linear regulator controller that acts almost as an active filter. This product can take a 1.2-V rail with a significant amount of voltage ripple on it and reduce the ripple to an acceptable level while offering fast enough transient response to meet the demands of the processor.
Fig. 3 shows a typical schematic for the implementation of the MIC5190 post-regulating Micrel's MIC2169 synchronous dc-dc converter to provide a 1-V, low-ripple voltage rail for the core voltage of a processor. The data sheet for the MIC5190 lists values for support components. The waveforms illustrated in Fig. 4a show the output ripple of Micrel's MIC2169, which is the input voltage to the MIC5190, along with the final output voltage ripple of the circuit, reduced by the MIC5190 from the 100 mV+ of the MIC2169 to less than 10 mV for the MIC5190 output. Fig. 4b shows the response of a commercially available dual-phase switcher under similar load conditions. The output ripple is greater than the post-regulated solution and the transient response is much worse.
Micrel's MIC5190 was specially designed to work with low RDS(ON) n-channel MOSFETs to offer high performance with a very low dropout voltage characteristics. The n-type pass transistor also gives very low output impedance, which allows for very fast response to sudden load changes. Fig. 4a also shows the load transient response for a 10-A load step from a 1-A nominal current. The amount of voltage droop is minimal thanks to the fast transient response of the loop.
Other tests have shown a switching regulator post-regulated with the MIC5190 had a transient response about twice as fast as that of a dual-phase dc-dc converter. Yet the 1700 µF total output capacitance in the dual-phase regulator was more than double the 770 µF of the output capacitor of the post-regulated dc-dc converter. Thus, the post-regulated converter can improve performance by a factor of four with the same capacitance, or the same performance can be achieved with one-fourth of the capacitance. The latter option can be especially useful if cost or form-factor reduction is the variable driving the design, as has been the case with some telecom applications.
Another advantage of post-regulated converters lies in their physical implementation on a PCB. Unlike the identical subcircuits in a dual-phase regulator, a post-regulated converter's dc-dc converter stage can be placed apart from the linear regulator, which is still placed at the point of load. Given the highly constrained design environment of PCB layout, this can have significant advantages.
In summary, existing solutions for powering low-voltage cores of networking processors or FPGAs that require high current are characterized either by a dc-dc converter with additional filtering stages in order to reduce ripple and improve transient response or by multiphase dc-dc converters. With a standard dc-dc converter and a high-speed linear regulator acting as a post-regulating filter, the power supply can be made into a true point-of-load regulator and offer the best-possible performance for the application.