Powering High-Performance ASICs and Microprocessors
Abstract: Today's highest performing ASICs and microprocessors can consume greater than 150W. With supply voltages of 1V to 1.5V, the required current for these devices can easily exceed 100A. By using multiphase DC-DC converters, the task of providing power to these devices is made more manageable.
Currently, scalable power-supply controllers are available that allow the designer to choose the number of phases for specific DC-DC converters. Scalability allows several controllers to be paralleled and synchronized. On-board PLL-based clock generators allow the controllers to be synchronized.
While no hard-and-fast power limit exists for a singlephase
buck regulator, the advantages of designing with
multiphase converters become apparent as load currents
rise above 20A to 30A. These advantages include:
reduced input-ripple current, substantially decreasing the
number of input capacitors; reduced output-ripple voltage
due to an effective multiplication of the ripple frequency;
reduced component temperature achieved by distributing
the losses over more components; and reduced-height
Multiphase converters are essentially multiple buck regulators
operated in parallel with their switching frequencies
synchronized and phase shifted by 360/n degrees, where n
identifies each phase. Paralleling converters makes output
regulation slightly more complex. This problem is easily
solved with a current-mode control IC that regulates each
inductor current in addition to the output voltage.
The key issue designers face when selecting input capacitors
is input-ripple-current handling. Input-ripple current is
substantially reduced by using a multiphase topology—the
input capacitor of each phase conducts a lower amplitude
input-current pulse. Also, phase shifting increases the
effective duty cycle of the current waveform, which results
in a lower RMS ripple current. The ripple-current levels
shown in Table 1 demonstrate the ripple-current reduction
and the input-capacitor savings.
High-k dielectric ceramic capacitors provide the best
ripple-current handling and the smallest PCB footprint.
Ceramic devices housed in an 1812 form factor exhibit
ripple-current ratings of 2A to 3A per capacitor. Electrolytic
capacitors are a good choice for cost-sensitive designs.
Accuracy requirements of <2% are commonly required for
core voltage supplies. For a 1.2V supply, this translates
to a ±25mV output-voltage window. A technique for using
the output-voltage window more effectively is called active
voltage positioning. At light loads, the converter regulates
the output voltage above the midpoint of the outputvoltage
window and, at heavy loads, regulates the output
voltage below the midpoint of the output-voltage window.
In the case of a ±25mV window, regulating at the high
end (low end) of the range during light loads (heavy
loads) allows the entire output-voltage window to be
used for a step-load increase (decrease).
Large load-current steps require both extremely low-ESR
capacitors to minimize transients and large enough capacitance
to absorb the stored energy of the main inductor
during a step-load decrease. Organic polymer chemistries
have improved low-ESR tantalum capacitors. Polymer
capacitors provide the most capacitance with the lowest
ESR. Ceramic capacitors have excellent high-frequency
characteristics, but the total capacitance per device is onehalf
to one-quarter that of tantalum and polymer capacitors.
Typically, ceramic capacitors are, therefore, not the
best choice as output capacitors.
A 12V to 1.2V converter requires 90% on-time from a
low-side MOSFET; conduction losses dominate switching
losses in this case. For this reason, two or three MOSFETs
are often paralleled. Operating several MOSFETs in
parallel effectively reduces RDS(ON) and thus lowers
conduction losses. When the MOSFET is turned off,
inductor current continues to flow through the MOSFET's
body diode. Under this condition, the MOSFET drain
voltage is essentially zero, reducing switching losses
substantially. Table 1 shows the losses for several multiphase
configurations. Note that the low-side MOSFET's
total losses decrease as the number of phases increases,
thus reducing the MOSFET's temperature rise.
With a duty cycle of 10 percent, high-side MOSFETswitching
losses dominate conduction losses. Because the
high-side MOSFET conducts for a small percentage of
time, conduction losses are less significant. Thus, low onresistance
is not as important as low switching losses.
During the switching intervals (both tON and tOFF), the
MOSFET has to withstand voltage and conduct current.
The product of this voltage and current determines the
MOSFET peak-power dissipation; therefore, the shorter
the switching interval, the lower the power dissipation.
When selecting a high-side MOSFET, choose a MOSFET
with low gate charge and gate-drain capacitance, both of
which are more important than low on-resistance. Table 1
illustrates how the total MOSFET losses decrease as the
number of phases increases.
The inductor value determines the peak-to-peak ripple
current. Allowable ripple current is typically calculated as
a percentage of maximum DC-output current. In most
applications, an optional ripple current is 20% to 40% of
the maximum DC-output current.
At low core voltages, the inductor current cannot decrease
as quickly as it can increase. During a load decrease, the
output capacitor can be overcharged, causing an overvoltage
condition. By using an inductor of smaller value
(allowing higher ripple current—closer to 40%), a lower
amount of stored energy is transferred to the output
capacitor, which minimizes voltage surge.
Table 1 provides estimates of heat-sinking requirements
for the number of phases used. In a forced-convection
cooling system that can provide 100LFM to 200LFM, a
single-phase design would require a fairly large heatsink
to achieve a 0.6°C/W thermal resistance. In the four-phase
design, the thermal resistance can increase to 2°C/W. This
thermal resistance is easily achieved without a heatsink
and 100LFM to 200LFM airflow.
Table 1. Comparison of critical parameters and the number of phases used for the design of
synchronous buck regulators. Example is a 12V to 1.2V, 100A buck regulator.
|| Number of Phases
| Current per phase
|Input capacitor, 3A rated
|RMS ripple current
|Power dissipation (each)
|Total power dissipation
|RMS ripple current (each)
|Power dissipation (each)
|Total power dissipation
|COUT 470µF, 10m
Figure 1 illustrates the MAX5038 configured as a fourphase
DC-DC converter. The MAX5038 master remotevoltage-
sense input (VSP to VSN pins) provides a signal
(DIFF) to both the master and the slave EAN inputs,
enabling parallel operation. The MAX5038 master also
provides a clock (CLKOUT) to the MAX5038 slave
controller. By floating the PHASE pin, the slave locks on
to the CLKIN signal with a 90° phase shift. The error
amplifier also performs the active voltage-positioning
function by setting the gain of the voltage-error amplifier.
Using precision gain-setting resistors ensures accurate
load sharing. The output of the voltage-error amplifier
(EAOUT) programs the load current of each phase.
Compensation (not shown) is provided for each current
loop at the CLP1 and CLP2 pins, providing a very stable
output for most line and load conditions.
Figure 1. A four-phase example using two MAX5038s. The master performs the remote voltage-sense function and the clock-generation function, which the slave controller uses to increase output current and synchronize the operating frequency.
Multiphase synchronous DC-DC converters effectively
power ASICs and processors that require 1V to 1.5V at
100A or more. They solve basic problems involving
capacitor ripple current, MOSFET power dissipation,
transient response, and allowable output-ripple voltage.
||Dual-Phase, Parallelable, Average Current-Mode Controllers
The content on this webpage is protected by copyright laws of the United States and of foreign countries. For requests to copy this content, contact us.
|© Jul 08, 2004, Maxim Integrated Products, Inc.
APP 3177: Jul 08, 2004
APPLICATION NOTE 3177,