- Buyers Guide
The III-V vs. Silicon Battle
Power amplifiers are important components in almost all wireless communication systems. They normally consume large amounts of power, and therefore play a critical role in battery life for mobile devices. As a rough estimate, in a typical WiMAX radio, the baseband and transceiver will consume about 600 mW, whereas the power amplifier will consume about 1.3 W.
When designing a power amplifier, there are a large number of options to be considered. One fundamental choice, however, is whether to use Silicon or III-V technology. This article will point out a number of important issues that affect power amplifier design, and will discuss advantages and disadvantages of the various underlying semiconductor technologies in determining who wins the III-V versus Silicon battle.
In recent years, there have been a number of technological changes that have had an impact on power amplifier designs. Technologies that use OFDM, like WiFi, WiMAX and LTE, are probably the most challenging for a power amplifier; they require a high degree of linearity to meet the required SNR targets, but must also handle a large peak-to-average power ratio associated with OFDM.
In addition, the 802.11-based WiFi and 802.16-based WiMAX standards have become some of the fastest growing technologies in use today, so it makes sense to focus on the GaAs versus Silicon debate within the context of low power (<1 W), high linearity OFDM power amplifiers.
Having chosen GaAs or Silicon, the power amplifier designer is then confronted with further options within each technology, and each option has its own set of advantages and disadvantages. In GaAs, one can design with GaAs HBT (bipolar-based), GaAs PHEMT (FET-based), or GaAs BiFET (a mixture of both bipolar and FET technologies).
In Silicon, one can design in CMOS (FET-based), or in higher speed SiGe BiCMOS (a mixture of both bipolar and FET technologies). The main workhorses in OFDM power amplifier design today are GaAs HBT and SiGe BiCMOS. However, CMOS as well as GaAs BiFET and PHEMT devices are also all in use.
To deliver high power with OFDM, GaAs has almost always been used due to a better trade-off between transition frequency, Ft, and breakdown voltage. However, over the past 10 years, Silicon technology has developed to the point where it is becoming harder to choose one technology over the other. A few years ago, anything above 2 GHz and/or 50 mW would have been designed in GaAs. Today, SiGe BiCMOS power amplifiers can be used at power levels close to 1 W and they have plenty of available gain even at 10 GHz.
If efficiency is important, GaAs technologies still offer the best performance, especially at higher powers. GaAs technology also offers higher breakdown voltage, which translates into greater robustness. However, circuitry has been developed that can protect lower breakdown Silicon devices. Complicating the picture even more, integrated CMOS PAs are now being considered at 2.4 and 5 GHz in applications where lower output powers (less than 15 dBm) and relatively low efficiencies (about 10 percent) can be tolerated.
Today, many wireless communications technologies exist, and they often operate simultaneously. For example, Bluetooth and cellular radios must both operate when using a Bluetooth headset during a voice call. WiMAX and cellular radios will both be active on mobile devices during handovers from one network to the other. Cellular and GPS radios will both be enabled when GPS is used on a cellular phone.
Figure 1 Dual-mode W-CDMA/WiMAX radio.
Coexistence refers to the simultaneous operation of multiple radios within a device. Figure 1 shows a typical example of a dual-mode WiMAX/W-CDMA radio. In this example, a W-CDMA daughter card is placed on top of, and in close proximity to, a WiMAX module. If the WiMAX and W-CDMA radio must operate simultaneously (which would be required during a handover from one network to another), then care must be taken to ensure that the radios do not interfere with one another.
But what does this have to do with power amplifiers? Since W-CDMA and WiMAX radios operate on different frequencies, one might naively expect no issues when both radios are operating at the same time. The problem, of course, is that noise from one radio that is emitted in the passband of the other radio cannot be filtered out at the receiver, and this noise can desensitize the victim receiver. This problem is most severe when two radios are collocated in the same device, as Figure 1 illustrates, since signals from one radio arrive virtually unattenuated at the receiver of the other radio.
An example is useful to illustrate this problem. Consider a WiMAX radio operating from 2.5 to 2.7 GHz transmitting at 23 dBm, while a victim W-CDMA radio is attempting to receive a signal at 2.17 GHz. The task is to determine what the maximum noise level is that the W-CDMA radio can tolerate so that its sensitivity (i.e. the smallest signal it can detect) is degraded by less than 0.1 dB when the WiMAX radio is operating.
W-CDMA has a 3.84 MHz channel bandwidth and the standard requires a sensitivity of -117 dBm for a coded CDMA signal. Assuming a 21 dB coding gain (128 chip code length), the sensitivity will be -96 dBm/3.84 MHz, or -161.8 dBm/Hz. Based on this, the noise at the W-CDMA antenna would need to be below -170.9 dBm/Hz to result in 0.1 dB degradation in sensitivity (-178.1 dBm + -161.8 dBm results in a net noise of -161.7 dBm).
Of course, the noise power emitted from the WiMAX PA will be reduced as the signal travels from the WiMAX Tx antenna to the W-CDMA Rx antenna. Since the two radios are located very close together, however, one can only expect approximately 20 dB isolation between the antennas, so the output noise from the WiMAX radio will need to be below -150.9 dBm/Hz.
Now that the output noise target for the WiMAX radio has been calculated, consider the implications on the power amplifier. Suppose that the input noise to the power amplifier is at the noise floor (-174 dBm/Hz), that the PA has a gain of 30 dB at 2.17 GHz, and has a noise figure of 5 dB. Therefore, the net noise from the PA will be -174 + 30 + 5 = -139 dBm/Hz, requiring 12 dB additional filtering at 2170 MHz in order to degrade sensitivity of the W-CDMA PA by only 0.1 dB.
The most obvious place to put the filter is directly after the PA. This is a bad choice, however, since any losses after the PA result in significant additional power consumption, and this power consumption is manifested as heat that must be dissipated. In addition, the effect of the filter loss is worse as output powers increase. For example, assuming that the coexistence filter has 1.5 dB loss and that the PA has 20 percent efficiency, Table 1 shows the effect of this filter on power consumption and net PA efficiency for different output powers. At an output power of 18 dBm, the 1.5 dB filter loss results in about 130 mW extra DC power consumption. Some of this power is dissipated in the filter (26 mW), but most of the additional power (104 mW) is dissipated in the PA, which must be made 1.5 dB larger to overcome the filter losses.
Table 1 Effect of Post PA Losses on Power Consumption
When transmitting at 23 dBm, adding the filter increases power consumption by 411 mW. At 26 dBm, the power consumption increases by 821 mW. One can see that putting a filter after the PA can result in a severe energy penalty (especially at higher output powers), and this results in shorter battery life. In addition, there are resultant cost increases since the PA must be made larger to overcome the filter loss. It is also interesting to note that a 1.5 dB post-PA loss reduces the PA efficiency by the same amount at each output power, from 20 to about 14.2 percent.
In order to reduce power losses, it is preferable to not place the coexistence filter after the PA. However, it should also not be placed in front of the PA, since most of the noise is generated within the PA. Therefore, the filter is optimally placed between the PA stages, internally on the PA die. The next question then is which technologies are best suited for implementing integrated filters?
At first glance, one might expect GaAs-based semiconductor technology to have an advantage because the passives have higher Q due to lower substrate losses. However, Silicon processes have evolved, and it is now possible to fabricate passive devices on insulating SiO2, and their performance can be as good as it is on lower-loss GaAs substrates.
There is an additional consideration, however. Modern foundry production tolerances make it very difficult to control the capacitance and inductance of passive devices to the accuracy required for demanding coexistence filters. In order to meet coexistence noise requirements, some form of post-production tuning is required. It is much more convenient to tune devices if one has access to digital control lines. The ability to integrate analog or digital control in tuning sharp filters in SiGe BiCMOS or Si-CMOS technology gives Silicon technology an advantage in this area.
Moore’s Law is bringing down the price of digital hardware, and this makes digital adaptive predistortion (DAPD) more attractive every year. In a DAPD system, shown in Figure 2, the output from the power amplifier is sampled, downconverted to baseband, and is then compared with the input signal. Phase and amplitude distortion created by the power amplifier are detected, and the baseband signal is adjusted to exactly counteract these distortions. This technique can be used to improve the overall PA efficiency.
Figure 2 Block diagram of a digital adaptive predistortion system.
Predistortion comes at a cost, however. Additional power is required to downconvert the RF output signal and to carry out the appropriate signal processing. One must always ensure that the improvement in efficiency outweighs the cost of implementing the additional functionality. However, DAPD overhead is typically fairly low, since updates to the lookup table can occur quite infrequently and the DAPD blocks are powered off most of the time.
Where does DAPD have the biggest impact? It will probably have the biggest impact when used with PAs developed in highly nonlinear processes. It also has the most significant impact on larger amplifiers, where the power required by the predistorter is dwarfed by the power used by the amplifier.
For example, integrated CMOS PAs are now being seen in low power WiFi handset devices. These CMOS PAs have very low Ft, and would need to operate at very high current densities to achieve the linearity required to meet WiFi EVM specifications.
When these devices are operated at lower current levels, they become very nonlinear and DAPD is a necessity for WiFi devices that use integrated CMOS PAs. Even with predistortion, the efficiency of integrated CMOS PAs is quite low, typically less than 10 percent. However, since these devices are operating at relatively low output powers (typically less than 40 mW), the efficiency is not that critical, and DAPD is used to ensure adequate linearity.
In contrast to CMOS only devices, the linearity of GaAs and SiGe transistors reduces the need for predistortion. However, predistortion can be used to improve performance, as it can improve both EVM and spectral mask.
For optimal performance with DAPD, it is best to use a PA that has been designed for maximum efficiency and not maximum linearity. In addition, by optimizing the feedback, one can tune the predistortion to apply more correction to EVM or mask. This can be important. For example, as output power increases, WiFi PAs become significantly mask limited because the out-of-band emissions limits specify a maximum absolute emission level. However, other systems like Japan’s new xgPHS system employ 256QAM modulation, and one would want to optimize the DAPD for EVM correction.
There is not really a preferred technology for DAPD. Predistortion is not a necessity for GaAs or BiCMOS PAs, but it can certainly help, and will improve efficiency, especially at higher output powers. For CMOS PAs, predistortion is a requirement due to the relatively low efficiency of this technology.
Most often, power amplifiers are specified in terms of current consumption at their rated output power, and power added efficiency (PAE) is normally specified at full power. When the output power is reduced, the current drawn by the PA is reduced. However, the current drawn is not linear with output power. For example, if the output power drops by 50 percent (3 dB), the current typically drops by only about 20 percent. In addition, when output power is backed off so that it approaches zero, the current does not drop to zero, but instead saturates at the PA quiescent current, Icq, due to the bias currents drawn by the PA.
In many applications, quiescent current is of no concern at all. For example, if a power amplifier will be operating at close to maximum power whenever it is transmitting, the power it consumes when backed off is unimportant, and Icq is irrelevant. This is typically the case for 802.11 WiFi power amplifiers: When data is being transmitted, the PA is on and always operating at maximum power; between transmit bursts it is disabled and consumes only leakage current.
If a PA must be used over a wide range of transmit powers, then power consumption at backed off power levels becomes important, and Icq is important. A good example of this occurs in either CDMA or WiMAX power amplifiers. WiMAX, for example, requires a minimum of 45 dB transmit dynamic range, since power control is intrinsic to the overall network.
Figure 3 Transmit power distribution for devices in CDMA and WiMAX networks.
Figure 3 shows the expected transmit power distribution for a mobile device in both a CDMA and WiMAX network. For CDMA, one can see that the handset most often transmits at -4 dBm, and it transmits at maximum power relatively infrequently. For WiMAX, handsets will most often transmit at approximately 10 dBm, and again, devices will transmit at maximum power only infrequently.
Also overlaid on this graph is current consumption versus output power for a typical power amplifier. Because the PA is often transmitting at low powers, one can see that it is important to minimize current consumption at lower output powers in order to maximize battery life.
There is probably little advantage of one technology over another in terms of getting good efficiency at backed off powers; they are all equally bad. For example, typical WiMAX PAs have 100 mA Icq. Assuming that the PA draws Icq when delivering 0 dBm, the PA consumes 330 mW and has an efficiency of only 0.3 percent at 0 dBm output power versus an efficiency of about 20 percent at full rated power.
There are a number of techniques that can be used to reduce power consumption at low output powers. A common technique is to bypass the output stage at low output powers, routing the RF energy around the final stage. This drops the gain, and significantly reduces current draw, since the output stage draws no current when it is bypassed. This technique is effective since the output stage is the largest stage, and draws the most current.
Typically, output stage bypass is done with switches. This favors technologies that have FET switches, since FET switches have much lower loss and are more linear. Therefore, PHEMT or GaAs BiFET processes are good choices.
A SiGe BiCMOS process, at first glance, might not seem to be a great choice since the technology makes it difficult to produce high quality, low loss switches due to substrate coupling effects. However, Silicon-on-Insulator (SOI) technology has been developed in recent years, and SOI switches are now becoming available with performance rivaling GaAs switches. Therefore, a SiGe BiCMOS process is also suitable for developing low quiescent current devices.
It is much more difficult to fabricate efficient switches with GaAs HBT or CMOS technology, so these technologies are not suitable for output stage bypass commonly used to achieve ultra low quiescent current.
In all wireless systems, if there is no data to transmit, the PA is disabled and ideally it consumes no power at all. However, unless a switch is placed in series with the supply voltage driving the PA (which is not attractive because of cost, size and power consumption), the power amplifier will always have a supply voltage applied to the collector (bipolar devices) or drain (FET devices).
While the PA can be ‘turned off’, in practice there is always a small amount of leakage current that flows even when the PA is disabled. This leakage current is a parasitic battery drain, and reduces standby times for mobile devices. Low leakage is often specified as a firm requirement in devices like handsets where standby time is important.
The requirement for low leakage is met with most technologies. GaAs HBT, SiGe HBT and CMOS power amplifiers can all be manufactured with low leakage currents, typically under 10 μA. The one technology that may have a problem with leakage current is PHEMT. These devices typically have leakage currents an order of magnitude larger than those manufactured with other technologies. The high leakage current seen with PHEMT PAs is intrinsic to this technology.
Technically speaking, a PHEMT gate looks like a diode, so the threshold voltage needs to be quite low (significantly less than a diode drop). As a result, with 0 V applied to the gate there can be appreciable leakage. Other technologies have insulated gates so threshold voltages are higher and leakage currents are much smaller.
The high leakage current of PHEMT devices is often cited as a reason not to use PHEMT technology for mobile devices. A PHEMT PA turned off and consuming a 100 μA leakage current would deplete a typical 1,000 mA-hr battery in 10,000 hours (417 days), and will have a minor impact on the mobile device’s standby time. While this seems to be a very small contributor, there are a large number of parasitic drains on the battery that all reduce standby time, and phone manufacturers wish to minimize each contributor.
So, for leakage current, the loser appears to be GaAs PHEMT. This is a factor in devices like mobile phones where standby time is important, but will be much less important in devices like laptops.
As Smartphones incorporating dual-band WiFi, multi-band cellular, GPS, FM and Bluetooth radios grow in popularity, it becomes increasingly difficult to fit everything into the required form factor. The RF front-end, comprising all components between the transceiver and antenna, can contribute significantly to the overall footprint. RF front-end vendors have responded, and the size of RF front-end components in communications devices has been continually shrinking.
Figure 4 Evolution of RF front-end sizes for WiFi radios.
Figure 4 shows a timeline giving an example of the degree to which integration has been applied to RF front-ends for WiFi, and one can see that integration has significantly reduced their footprint. In 2002, front-ends comprised unmatched PAs with many discrete devices, and the RF front-end had a size of about 16 x 18 mm. By 2005, front-end laminate-based modules were available incorporating discrete surface-mount components for matching, and the size had been reduced to about 8 x 7 mm. In 2007, many of these discrete matching components had been replaced by integrated passive devices, and one could now achieve the same functionality in a 4 x 4 mm module without the need for a laminate.
The next logical step in this integration process is to develop a front-end integrated circuit (FEIC), shown in the last photo in Figure 4, achieving a 3 x 3 mm form factor. FEICs offer the possibility of much greater levels of integration by integrating PAs, LNAs, switches and filters onto a single chip. Of course, the pattern of progressive integration has been repeated numerous times in the history of Silicon IC development. GaAs PHEMT and BiFET technologies are well suited for FEICs as they can be used to make excellent LNAs, PAs and switches.
As has been discussed, the SiGe BiCMOS process, at first glance, might not seem to be a great choice, since it is difficult to produce high quality, low loss switches with this technology. However, SOI switches are now available with performance rivaling GaAs switches. As a result, a SiGe BiCMOS process is also a highly suitable platform for FEIC development and one would expect significant growth in this area. In fact, the SiGe BiCMOS platform is even more compelling when considering the possible integration of battery management circuits onto the same die.
To summarize, for front-end IC development, CMOS and GaAs HBTs will not be suitable. GaAs PHEMT and BiFET processes, as well as SiGe BiCMOS processes incorporating SOI technology, are all good choices.
Power amplifiers with Serial Interfaces
Historically, PAs have been standalone, independent components. Even today, most PAs are controlled with only a single analog enable signal, often requiring precision regulators. In RF front-end modules where power amplifiers, low noise amplifiers and switches are all integrated into a single packaged device, routing the control signals from the baseband chip to the RF module can be very challenging, especially with the advent of multi-band and multi-PA MIMO technologies. For example, an 802.11a/b/g MIMO radio will require two 5 GHz PAs, two 5 GHz LNAs, two 2.4 GHz PAs, two 2.4 GHz LNAs, filters and Rx/Tx switches, each of which must be controlled separately.
A new trend that is emerging is to use a serial interface to control the PA and/or components within the RF front-end module. A serial-interface-controlled PA has the potential to revolutionize PA operation, bringing the digital interface one step closer to the antenna. In the context of complex front-end modules, the serial interface can reduce or eliminate control lines, greatly simplifying routing from the baseband chip. One could also use the serial interface to report temperature and detector voltages directly over the serial bus, thereby reducing pin-count and eliminating the need for A/D converters on the baseband chip.
Serial interfaces favor Silicon processes like CMOS and SiGe BiCMOS. Most GaAs processes lack complimentary devices (pFET or PNP transistors). As a result, it is not possible to implement significant logic or logic control like a truth table on a GaAs die. Therefore, HBT, BiFET, or PHEMT-based devices would all require an external CMOS logic die to properly implement a serial interface. Consequently, if serial interface control of PAs or RF front-ends is important, the logical choice is CMOS or SiGe BiCMOS.
There have been a number of important issues that have impacted the design of power amplifiers in recent years. This article has summarized several new issues, and has looked at how each affects the choice of technology for the power amplifier, particularly for PAs used with OFDM modulations. CMOS PAs are suitable for lower output powers, and require the use of digital adaptive predistortion to achieve linearity required for operation.
While GaAs HBT technology has traditionally been used for high power and high frequency power amplifiers, high performance SiGe BiCMOS power amplifiers are now competing very effectively with them. SiGe BiCMOS power amplifiers can be preferred to GaAs HBT PAs based on the availability of digital logic for serial interface control, as well as for the high levels of integration possible for front-end IC development. Consequently, GaAs HBT and GaAs PHEMT PAs will be used at progressively higher power levels and in more specialized applications. Slowly but surely, Silicon is progressing in the III-V versus Silicon battle on the power amplifier front.
Darcy Poulin holds a BS degree with honors in engineering physics from Queen’s University at Kingston, and a PhD degree in applied physics from McMaster University in Hamilton, Ontario, Canada. He brings to SiGe Semiconductor more than 15 years of experience in RF engineering and IC design. He is currently principal engineer, RF Systems and Technical Marketing, and is responsible for RF systems work, standards development, and technical marketing activities for WiFi, WiMAX and LTE.