Editor's Note. To view a PDF version of this article, Click Here.
Although often considered an afterthought, the switching power supply for a digital subscriber line (DSL) modem can have a seriously detrimental effect on the data throughput of that modem if care isn't taken up front with respect to the choice of switching frequency. Data loss is but one side-effect of a poor design; by following some basic design rules the designer can also make strides in overall modem cost reduction, efficiency and reliability.
DSL modems send their signals through a twisted-pair telephone line, which attenuates the high-frequency signal as the distance increases, causing the modem bit rate to start decreasing after just a few thousand feet. It continues to decrease with loop length, out to a maximum loop length of about 3 miles, where the connection becomes marginal.
Data rates as high as 8 Mbits/second are achievable with short loops but drop off to around 100 kbits/s with the longest loops. As the line length exceeds 3 miles, the bandwidth decreases, and data throughput drops off quickly. Eighty to 90 dB of transmission signal attenuation is possible at these line lengths, forcing the modem to implement noise-reduction techniques that produce high signal-to-noise ratios (SNRs), minimize crosstalk and interference, and have large dynamic signal ranges.
Unfortunately, the SNR of the modem has a trough at the power supply's switching frequency. That trough negatively affects the number of data bits transmitted, reducing the modem's overall data transfer rate.
To resolve the power supply-vs.-modem conflict, the designer must first examine the relationship between the power supply switching frequency and the modem's operating characteristics.
An asymmetric DSL (ADSL) modem uses a 25-kHz to 1.1-MHz frequency spectrum for data transmission. The downstream direction (to the client) uses the 138-kHz to 1.1-MHz band. The upstream direction (client to server) uses the 25-kHz to 138-kHz band. The modem divides that spectrum into 256 equally spaced carrier bands, or "bins," each having a 4.3125-kHz bandwidth. The carrier bands are allocated to voice transmission, upstream and downstream data, and buffer zones. The number of bits assigned to each carrier band depends on the level of noise and interference present.
Figure 1 shows how the switching noise from a power supply affects the number of data bits per carrier band on an ADSL client modem. An 80-mVpp, 250-kHz square wave imposed on the 3.3-Vdc voltage significantly reduces the available bits at the 250-kHz-carrier band, its 750-kHz harmonic and, to a lesser extent, the 500-kHz harmonic (because of the distortion present on the square-wave signal). The harmonics of a square wave follow a sin(x)/x function, which has peaks at multiples of the odd harmonics. The data in Figure 1 represents 11,000 feet of transmission line length.
Figure 2 shows the reduction in the overall modem bit rate vs. power supply switching frequency and ripple amplitudes. The data shows that both the amplitude and the frequency of the ripple affect the overall modem data rate. Increasing the ripple voltage amplitude reduces the overall modem data rate by reducing the SNR in the carrier bands corresponding to the ripple frequency and its harmonics.
Power supply vs. modem
Higher power supply switching frequencies have less of an effect on the overall data rate, for two reasons. The first is that higher switching frequencies affect higher-frequency bins, which have lower bit rates to start with. The second is that the harmonics are spaced farther apart, so fewer bins are affected.
Power supply switching frequencies greater than 1.1 MHz typically have a minimal or no impact on the modem data rate, since the fundamental of the switching frequency is higher than the frequencies used to transmit data. Reducing the switching frequency below 100 kHz will move the fundamental below the first downstream carrier band, but the harmonics will affect many carrier bands up to 1.1 MHz.
Other frequency effects
Based on throughput only, a power supply switching frequency of greater than 1 MHz is obviously desirable from an interference standpoint. But there are other power supply factors that are affected by the switching frequency. They include the supply's overall volume, cost, efficiency and reliability. None of those factors can be optimized independently of the others, and in many cases, optimizing only one supply parameter leads to a less-than-optimal overall solution.
Increasing the switching frequency of an ADSL power supply reduces the volume for two reasons: smaller passives and higher allowable ripple. Figure 2 shows that lower switching frequencies require a lower output ripple to maintain a data rate of 6.7 Mbits/s. For example, at 100 kHz, the allowable ripple for 6.7 Mbits/s is 10 mVpp, while at 1 MHz, the allowable ripple is 80 mVpp.
But there's a catch. Not only is more filtering needed at lower frequencies, but the size of those filtering components increases, since their size is inversely proportional to the switching frequency. For example, the 100-kHz circuit requires a 22-microHenry inductor; the 1-MHz circuit only requires a 2.2-microHenry inductor. The latter is much smaller since inductor size decreases with its value.
But the decrease in inductor volume with size is not linear. At some point, an additional increase in switching frequency does not significantly decrease the inductor volume. Also, as the inductor becomes smaller, the packaging and solder pads take up an increasing percentage of overall volume, thereby diminishing the overall reduction in inductor volume with increasing frequency.
To prevent coupling, shielded inductors, with closed flux paths, are preferred over "open," bobbin-type cores. The closed-core structures contain or hold most of the flux lines within the magnetic material, thereby preventing coupling into nearby components and etch. The flux field is highest near the gap of the inductor, so care must be taken during layout to keep sensitive and high-impedance circuits away from it. Shielded inductors have current ratings that are almost 25 percent lower than the nonshielded variety because of the additional core volume and the smaller-gauge wire that must be used.
The minimum input- and output-capacitance requirements are also inversely proportional to frequency. As the frequency increases, the required filtering is reduced, so lower capacitance values are required. At lower frequencies, ceramic capacitors do not have the bulk capacitance required for filtering; therefore, aluminum electrolytic, organic electrolytic or specialty polymer capacitors are used. Because those capacitors are ESR- and ripple-current-limited, their volume varies little with switching frequency.
At around 500 kHz and above, ceramic capacitors become a viable option since they have very low ESR and therefore are capacitance-limited. As the switching frequency increases, the amount of charge per cycle delivered by the capacitors is reduced. The capacitance values are determined by the charge requirements. Thus, the capacitance requirements decrease as frequency increases. However, as with inductors, packaging issues diminish the gains of reduced capacitor volume.
Although filter component areas are reduced with increasing frequency, the control circuitry area stays constant.
shows the overall power supply area as well as the breakdown of area between the control circuit and the filter. Note that the total area is reduced by about 23 percent from 100 kHz to 250 kHz but is reduced by only 13 percent from 250 kHz to 500 kHz.
Frequency vs. efficiency
Switching frequency has a direct impact on power supply efficiency. The main power loss component in a power supply is typically the power FETs. Power-FET losses comprise several terms, some of which are frequency-dependent.
Conduction losses in the FET are independent of frequency. This loss term is a function of power supply current and FET resistance. FET losses that are directly proportional to frequency, however, include gate drive, Coss (FET output capacitance), reverse-recovery, body-diode conduction and switching losses. Each of these terms dissipates a fixed amount of energy every time the power supply completes a switching cycle. The more times per second the power supply switches, the greater the power that is dissipated by the frequency-dependent loss components.
At lower frequencies, FET conduction losses dominate. At higher frequencies, the conduction losses stay constant while the frequency-dependent losses increase. At maximum output power, the fixed losses in the power supply are 892 mW at both 100 kHz and 1 MHz. The frequency-dependent losses are 165 mW at 100 kHz but increase to 1650 mW at 1 MHz.
Power supply reliability and cost are indirectly related to switching frequency. Reduced efficiency at higher switching frequencies translates into higher operating temperatures. The failure rate for semiconductors doubles for every 10 degrees C rise in temperature. An estimate of temperature rise above ambient is shown in Figure 4. The data assumes natural convection cooling with the board area of Figure 3.
Fortunately, the designer has several options when it comes to maintaining a reasonable temperature rise for the higher switching frequencies. The first is to increase the available area for cooling. That may include using either larger or multiple components in parallel, in conjunction with a larger PWB to help spread the heat. For example, replacing one SO-8 package with two SO-8s or a D-Pak can significantly reduce junction temperatures. The second option is to provide additional cooling in the form of fans, heat sinks or both. Both options increase the overall cost and size of the supply. Power supplies are in production that cover more than 2.5 square inches of surface area, running with a temperature rise of only 45 degrees C at 1 MHz and maximum load.
In the end, it's a constant trade-off of switching frequency vs. ripple, size and efficiency while maintaining a constant modem throughput. The power supply designer must act as a system engineer at the power supply level, taking into account all design parameters in order to produce the optimum solution.
If minimum area is a concern, the designer may choose a switching frequency of around 500 kHz. At 500 kHz, the power supply area is near a minimum because of the decreased sensitivity of the modem to switching noise, and the efficiency is still at an acceptable level.
If minimum losses are the driving factor, the supply should be run somewhere between 200 kHz and 300 kHz. Here, frequency-dependent losses are kept to a minimum, and the power supply efficiency is near its maximum level. But the volume is increased because of additional filtering requirements to maintain an acceptable modem throughput.
- GCI Technologies, Plano Texas
- Kollman, Robert, and John Betten, Power Supplies for Residential Telephony Systems, Texas Instruments.
About the Authors
John Betten is a senior design engineer and a member of the Group Technical Staff at Texas Instruments. He has a BSEE from the University of Pittsburgh and can be reached at firstname.lastname@example.org.
Michael Day is a power management application supervisor for portable power at Texas Instruments. He has a BSEE and MSEE from Texas Tech University (Lubbock) and can be reached at email@example.com.