August 15, 1997

Powering the big microprocessors

BILL TRAVIS,SENIOR TECHNICAL EDITOR

Driving the big microprocessors is no picnic; your power-supply design must furnish many amps at a tightly controlled voltage level and respond quickly to heavy load transients.

Time was, supplying power to a microprocessor was no big deal. You could feed it approximately 5V at a few hundred milliamps, and it would compute away happily at its laughable 4.77 MHz or so. Today, a power-supply designer's life is more difficult. Multimillion-transistor, multihundred-megahertz mPs consume many amps and place rigid demands on voltage levels. Everybody knows that a CMOS IC consumes no power--unless you use it. Granted, CMOS eats no "real" (static) power, but charging and discharging millions of tiny capacitors takes real current--and lots of it. IC manufacturers are responding to this need for current with a spate of dc/dc converters and buck-controller ICs that use PWM techniques to achieve high efficiency.

High efficiency is eminently desirable; it equates to less wasted power. But wasted power is not really the issue in computer systems; the real issue is heat. Each watt wasted in a power-supply design results in an incremental temperature rise. The mP itself may dissipate approximately 40W, and that's why some of the big processors come with a muffin fan and heat sink attached. To dissipate another 40W or so in the nearby power source would be intolerable and thermally unmanageable. Ergo, the use of a dc/dc converter is almost always mandatory. Imagine using a linear regulator to reduce a 12V line to 2.5V at 13A: 123.5W of power and heat would be wasted in the regulator.

All the big µPs use multiple power supplies: 3.3V for the I/O ports and internal glue logic and a lower voltage for the core. For example, the core voltage for a 150-MHz Intel (Santa Clara, CA) Pentium Pro is 3.1V. For a DEC (Hudson, MA) 21164 366-MHz Alpha chip, the core voltage is 2.5V. Cyrix's (Richardson, TX) 233-MHz PR266 uses a 2.8V supply for its core. The core is the big power consumer in a µP. The Intel Pentium Pro takes 9.9A maximum and 11A peak (for several microseconds). A Pentium II (Klamath) consumes currents as high as 14.2A. The DEC and Cyrix chips take 11 and 9.2A maximum, respectively.

Not only do the big µPs want a lot of current, but they also want it fast. Intel's P6 needs a supply that can furnish 10A in 10 nsec. Realizing that no power supply can deliv er that much current that fast, Intel's specifiers eased the requirement to 30A/µsec, a factor of 33 slower than the µP's requirement. It's the job of low-inductance, low-ESR, multilayer ceramic capacitors (MLCs) to provide a localized (right next to the processor's pins) current reservoir for the fast transients.

The buck regulator is the topology of overwhelming choice for stepping down the voltage for the big µPs (Figure 1a). V_CC is usually 12V, and V_IN can be 5 or 12V. 12V for V_CC and 5V for V_IN are a good combination, because the gate of the rectifying MOSFET needs a drive signal whose voltage is considerably higher than V_IN to fully turn on the device. However, V_CC can be equal to V_IN if you use a charge-pump structure to generate high gate voltages. The circuit in Figure 1a is a nonsynchronous configuration; it uses a Schottky "catch" diode to supply current to the inductor when the MOSFET turns off.

Buck regulators use PWM to generate and regulate the output voltage. If the MOSFET and catch diode were perfect switches, the whole regulator structure would dissipate no power (except for the regulator IC's few milliamps of operating current), thereby achieving 100% conversion efficiency. However, the MOSFET and diode are not perfect devices: The MOSFET has a finite R_DS(ON), and the diode has a forward drop. A synchronous structure reduces the losses by replacing the Schottky diode with another n-channel MOSFET (Figure 1b). Reference 1 calculates the power savings. If the diode's forward drop is 0.64V and the peak current is 10A, the peak power wasted is 6.4W. The average power is a function of the duty cycle of the output waveform. A MOSFET with 22-microhms R_DS(ON) reduces the wasted power to 2.2W.

The voltage-code lines in Figure 1 apply to Intel's Pentium and AMD's (Sunnyvale, CA) K6 µPs, which deliver a voltage-specifier code to the power supply. To squeeze the maximum available speed out of each processor, testing determines and digitally encodes the µP's optimum operating voltage. Other big µPs on the market work with a fixed voltage (for example, 2.5 or 2.8V). The input code drives a D/A converter in the regulator. A difference amplifier compares the output of the D/A converter with the fed-back output voltage, and a comparator adjusts the duty cycle of the PWM waveform to bring the output voltage to the desired value.

A number of digitally controlled buck-regulator ICs work with external components--MOSFETs, Schottky diodes, inductors, and capacitors--to configure power supplies that conform to the requirements of the various Pentiums and other µPs (Table 1). You can, of course, use any of the digitally controlled regulators in fixed-voltage applications by simply tying the appropriate voltage-code pins to ground or to V_CC. Note also that most of the manufacturers in Table 1 offer other fixed-voltage or adjustable (via an external resistor) regulators. One typical application configuration is for Cherry Semiconductor's CS-5150 buck controller (Figure 2).

The design uses a lot of capacitance: 4800 µF at the input and 6000 µF at the output. In a buck regulator, the current peaks and spikes at the output reflect directly back to the input. The 4800 µF is necessary to absorb the current transients and prevent them from invading the main power supply. At the output, the large amount of capacitance is necessary to keep the output voltage within specification during current transients. Reference 2 goes through some fairly involved calculations to determine the output capacitance needed to satisfy a 3.1V±5% output-voltage spec. The calculations assume a 1% reference accuracy, leaving ±4% (±125 mV) for regulation and transient overshoot.

Taking account of the ESR, ESL, and value ranges of various types of capacitors, the calculations yield the required series inductance and minimum switching frequency. Using aluminum electrolytic capacitors, you'd need five units totaling 7500 µF at a minimum switching frequency of 50 kHz. Or you'd need 11 chip tantalums, totaling 1100 µF, at a minimum frequency of 380 kHz. Eight MLCs, totaling 176 µF, would do the job at a switching frequency of 1 MHz. The MLC solution sounds like a good deal, but you must be aware of the trade-offs: MOSFET switching losses and inductor losses are a function of frequency and would be fairly onerous at 1 MHz.

The choice of capacitors for a high-current buck regulator is serious business. Contrary to what you'd expect, the capacitors on the input terminal have the most stringent requirements on ESR. Assuming a 50% PWM duty cycle, the rms ripple current in these capacitors equals half the output current. With these high current values, a high ESR in the input capacitors would produce excessive heating. You also need low ESR in the output capacitors, of course, to minimize voltage steps accruing from current transients. The peak-to-peak ripple current in the output capacitors ranges from 10 to 40% of the load current.

If you don't want an early failure in your computer design, you must carefully choose the input and output capacitors. You can use low-ESR aluminum electrolytics, solid tantalums, or a great bank of MLCs. Make certain your design doesn't exceed the ripple-current ratings of the electrolytics, and be sure you specify surge-tested tantalums. Linear Technology's LTC1553 data sheet states that "low-cost, generic tantalums are known to have very short lives followed by explosive deaths in switching power-supply applications." Note that all the recommended application circuits for the products described here use multiple capacitors in parallel. The parallel connection is not to augment the capacitance, but rather to minimize the ESR.

In Reference 3, Bob Dobkin, vice president of engineering at Linear Technology, discusses the degradation with age of aluminum electrolytic capacitors: In short, the electrolyte dries out with time, and the ESR increases. Note that the ripple-current ratings for the aluminum capacitors are usually valid for only 2000 hours of operation. Dobkin strongly recommends using solid-tantalum capacitors, which don't have a wearout mechanism. Other industry experts maintain that aluminum electrolytics are just fine if you specify them conservatively. They agree that degradation is in-evitable, but so is the obsolescence of new computers, which become dino-saurs in a very short period.

Code sets output voltage

In Figure 2, the 4-bit code from the µP sets the regulator's output voltage to a value of 2.1 to 3.5V in 100-mV increments. Cherry sets each voltage 40 mV higher (2.14 to 3.54 mV), which makes sense if you think about interconnect and socketing losses. For 5-bit regulators, Intel specifies voltage levels of 1.8 to 2.05V in 50-mV increments and 2.1 to 3.5V in 100-mV increments. Some of the controllers in Table 1 conform strictly to the Intel requirement and supply only those voltages; others use the non-Intel, "throwaway" codes to generate voltages as low as 1.3V. For example, Cherry's CS-5155 extends the low end down to 1.34V. These controllers pave the way for future lower supply voltages.

The Cherry regulators use "hiccup-mode" overcurrent protection. The scheme involves charging and discharging the "soft-start" capacitor connected to the SS terminal. In the event of a short circuit, the regulator shuts off the upper MOSFET and then keeps trying to restart at a 1.65% duty cycle. Cherry claims that the hiccup action results in less stress to the regulator components, input power supply, and pc-board traces than you'd obtain with constant-current limiting. In the event of an overvoltage condition, the regulator shuts off the top MOSFET and turns on the bottom one, thereby providing a "crowbar" to protect the load. The Cherry circuits also provide two feedback pins: V_FB ensures dc accuracy of the output voltage, and V_FFB fine-tunes the output in the presence of fast transients. The fine tuning takes 25 µsec for a 0 to 13A step.

A nonsynchronous buck controller from Cherry, the CS-5151, uses the same connections as those in Figure 2 but substitutes a Schottky rectifier for the bottom MOSFET. The nonsynchronous configuration sacrifices a few percentage points in efficiency but saves approximately $0.75 in materials cost. In the highly competitive PC market, this cost reduction is significant. This year, expect a 5-bit regulator from Cherry that incorporates a power-good indicator and an output-enable function. You can gain some efficiency by connecting a Schottky rectifier in parallel with the bottom MOSFET in the synchronous converter or by using two Schottky rectifiers in the nonsynchronous design. Another way to reduce losses is to use two MOSFETs in parallel for the high-side switch.

Charge it at the pump

Standard power MOSFETs need approximately 6V gate-source voltage to fully turn on; logic-level MOSFETs require fewer volts. Many of the controllers in Table 1 use a 12V supply, so with a 5V input to the upper MOSFET, the devices have no trouble providing full MOSFET enhancement. However, in a single-supply, 5 or 12V system, the high-side driver needs a separate supply or a charge-pump (bootstrap) arrangement to generate the needed gate-source voltage. Most of the controllers in Table 1 provide a separate high-side supply pin.

If the high-side driver in your controller has an independent supply pin, you can use one of the bootstrapping circuits (Figure 3), derived from Linear Technology's LTC1553 data sheet, to generate the necessary gate-source voltage for the upper MOSFET. The circuit in Figure 3a is a voltage doubler that's effective with logic-level MOSFETs but inadequate for standard MOSFETs. For high values of V_IN, the 1N5248B zener diode clamps the supply rail to 18V, 2V below the IC's abso-lute-maximum rating. The voltage tripler in Figure 3b provides adequate drive for standard MOSFETs. Figure 3c shows a 17V charge pump that uses both the 5 and the 12V supplies. Logic-level MOSFETs are the devices of choice for this circuit. Note that the LTC1553 and Motorola's (Phoenix) MC33470 (to be introduced this fall) are the only Pentium-class, multiple-sourced, pin-compatible, buck-controller ICs on the market.

In a synchronous converter, it's important to avoid cross-conduction, or simultaneous conduction of the upper and lower MOSFETs. Most of the synchronous controllers in Table 1 provide a break-before-make function to prevent this dangerous condition. Unitrode's UCC3882, for example, offers programmable dead times between the turn-off and -on of the two MOSFETs. In addition, overvoltage and undervoltage comparators in the UCC3882 monitor the system voltage and shut off the controller if the voltage goes outside a ±7.5% window. Unitrode also provides a reference voltage with better than ±1% accuracy in all its P6-compatible products. Given the Intel ±5% accuracy spec, this tight reference allows more room for transients.

Upping the ante in integration

The devices described so far all need external MOSFETs to configure a complete regulator. Two regulators from Elantec incorporate the MOSFET power switches. The EL7563C provides µP currents as high as 12.4A and uses on-chip, resistorless current sensing to effect current-mode control. The device uses bootstrapping to provide the gate-source drive for the high-side switch. Reverse lead bending allows you to directly solder the integral heat slug to a pc board. The lower power EL7556C also has on-chip power MOSFETs and supplies currents to 6A. It's not a code-programmable regulator; you set the output voltage with two external resistors.

If you don't mind connecting a few extra external components, some controllers other than the code-settable ones in Table 1 are also suitable for powering the big µPs. For example, the ILC6330 from Impala Linear Corp is a synchronous controller that switches external MOSFETs to supply currents as high as 14A. To make the controller compatible with µP voltage codes, all you need is a quad CMOS gate and some resistors (Figure 4). The device uses bootstrapping to configure a charge pump to drive the high-side MOSFET; consequently, the device can accommodate input voltages from 4.5 to 25V. The IC is also suitable for GTL⁺-bus termination, which requires 1.5V termination at both ends of the bus and 5.38A total current.

The L4990 primary controller from SGS-Thomson also uses external gates and other components to configure a voltage-regulator module (VRM) for the big µPs. The device is a nonsynchronous buck controller that uses current-mode control. It has programmable soft-start, a disable pin, and pulse-by-pulse current limiting. An overvoltage-protection circuit shuts off the IC if the output voltage exceeds 18% above the nominal value.

The earlier comments about linear regulators notwithstanding, in some cases, a low-dropout regulator can do an effective job of powering the big mPs. Linear Technology's LT1581, for example, specs a low 430-mV dropout voltage at 10A. It's designed to drop 5V to 3.xxV for Pentiums operating at greater than 90 MHz or to drop 3.3V to 2.9V for Portable Pentium µPs. The use of a linear regulator has multiple advantages in some applications. First, it eases the ripple and accuracy requirements on its driving dc/dc converter in a multistep-down configuration. Next, it needs much lower output capacitance (700 µF) than a straight switcher requires. Finally, it offers exceptional load regulation: ±1 mV (typical) from no load to full load. The LT1581 costs $7 (1000). Other Linear Technology products, the $2.40 (1000) LT1575 fixed regulator and the $4.45 (1000) LT1577 dual fixed/adjustable regulator (Figure 5), rely on an external n-channel MOSFET to achieve low dropout voltages.

VRMs and reference designs

Buck-controller ICs such as the ones described here are integral parts of complete VRMs, which are available from several manufacturers. VXI Electronics, for example, supplies 5-bit-programmable VRMs that comply with Intel's VRM 8.1 spec, Revision 1.4. They're available in 5 and 12V versions. The modules supply 0 to 14.5A in 350 nsec. Again, this speed is much too low for a blazing µP; a reservoir of local multilayer capacitors satisfies the µP's need for instant current. Astec America, Linfinity, Celestica, and TDK also supply complete VRMs that meet Intel specs. Astec is evidently conscientious about capacitor selection; the AA32/AA39 data sheet specs 500,000 hours MTBF at 55°C.

The PT770x "Big Hammer" 4- and 5-bit power modules from Power Trends operate from 5V and supply 2 to 3.5V at 15A. Various models cover 1.3 to 3.5V, including the Intel-specified 1.8 to 2.05V and 2.1 to 3.5V. The modules incorporate all VRM features except the output capacitors. Their relatively high 750-kHz switching frequency keeps the value of the needed output capacitors to 330 µF×4, modest by VRM standards. PT7749 current-booster modules provide additional output current in 15A increments. Both series come in a compact, 27-pin SIPs. The PT700x and PT7749 cost $29.50 and $27.50, respec tively.

If you prefer to design your own µP-class regulator, you have access to Intel-compliant reference designs from most of the IC manufacturers in Table 1, as well as from International Rectifier, a manufacturer of the low-R_DS(ON) MOSFETs and the Schottky diodes you find in the VRM's output section. The company's design uses 14-microhms IRL3103 "Super Fetky" MOSFETs in a 200-kHz, synchronous-rectifier configuration. The buck-controller IC is a Linear Technology LTC1430 with a simple CMOS-switch, 4-bit DAC in its feedback divider. The reference design comes with a bill of materials, complete specs, and a pc-board layout in Gerber format. A $50 design kit includes the design specs, an evaluation board, and the power MOSFETs.

Providing power to current-hungry µPs is not easy, and it's not going to get any easier. As operating voltages continue to lower, efficiency in the regulator inevitably decreases because of rectification losses. You thus have to pay close attention to effective heat sinking in the output section of the VRMs you design. The bright side is that the ready availability of feature-laden buck-controller ICs makes your job of controlling the power switches easy.

References

1. Mammano, Bob, "Fueling the megaprocessors--empowering dynamic energy management," Power-Supply Design Seminar Notes, Unitrode Integrated Circuits, 1996.
2. O'Connor, John, "Converter optimization for powering low-voltage, high-performance microprocessors," Unitrode Integrated Circuits, 1996.
3. Dobkin, Bob, "Alert! PCs built to fail," Linear Technology Corp, 1997.
4. Kanner, Martin, "Capacitor amplifier reduces ripple without dc loss," EDN, June 6, 1996, pg 137.

| EDN Access | Feedback | Table of Contents |