5G front-end module architectures

The smartphone market is very dynamic and the move to 5G requires coverage for more than 50 LTE bands from 500 MHz to 6 GHz [Reference Balteanu, Serge Drogi, Modi, Choi, Khesbak and Agarwal13]. These bands cover the entire world with very few mobile SKUs. A lot of research has been done for single die power amplifier (PA) in different technologies such as silicon germanium (SiGe), complementary metal-oxide semiconductor (CMOS), and silicon on insulator (SOI) [Reference Lie, Tsay, Hall, Nukala and Lopez14] but still gallium arsenide (GaAs) is the technology mainly used. With the band proliferation and the transition to 5G the weight in terms of cost and size is determined by the number and type of acoustic filters, such as surface acoustic wave (SAW) and bulk acoustic wave (BAW). The acoustic filters are placed with other components such as PAs, SOI switches, and low noise amplifiers (LNAs) into an FEM with acoustic duplexers/filters (FEMiD). With the increase of the number of antennas and the requirements such as each antenna to be reached by multiple PAs and FEMs, the RF path loss to the antennas, on the mobile device, becomes higher for 5G as presented in Fig. 7. The RF loss can be 4–5 dB, and this RF loss is higher than the typical loss for 3G/4G (3 dB). Unfortunately, this RF loss must be compensated with higher RF power delivered by the PA and increased SOI RF switch size to maintain the linearity.

Fig. 7. 5G RF path loss from PA output to the antenna output.

Sub 3 GHz bands provide primary LTE cellular coverage, and the new 5G bands provide the increased capacity using MIMO and CA. The typical FEM structure is presented in Fig. 8.

Fig. 8. LTE 4G/5G RF front-end module structure.

The legacy 2G/3G module is directly connected to the battery but the 4G/5G modules are connected through a power management integrated circuit used also for ET [Reference Paek, Kim, Bang, Baek, Choi, Nomiyama, Han, Choo, Youn, Park, Lee, Kim, Lee, Cho and Kang15]. The TDD FEM can use also the LNAs on the same module; due to isolation this is not a solution for FDD systems that have both UL and DL operating at the same time. This is one aspect of the FEM architecture which must be considered from the beginning together with other mitigations due to multiple intermodulation products for multiple RF transmitters and receivers operating at the same time. Intermodulation products are determined by all the RF Tx paths, including WiFi [Reference Maldonado, Karstensen, Pocovi, Esswie, Rosa, Alanen, Kasslin and Kolding16], and all other clock-related activity as well (charge pumps, MIPI SPI clocks, etc.). There are few types of interference due to simultaneous UL and DL over different bands in CA configurations which will degrade the Rx sensitivity (desense):

• • Interference from sub-harmonic mixing for CA case when the higher UL frequency signal is a multiple of the lower frequency (Band 7 and Band 27/CA case) and desense the LB Rx.

• • Interference from the harmonic of lower frequency UL signals to the higher frequency DL when the harmonic of UL lands into UL Rx frequency band. For example, when a UE is transmitting on Band 3 4G/LTE and receiving on 5G NR bands n77/n78. Second harmonic of Band 3 will land into 5G NR Rx for bands n77/n78. Another CA case for desense is when 3rd harmonic (H3) from LB lands in RX high band (HB) as presented in Fig. 9.

• • Intermodulation distortion (IMD) products between different Tx frequencies and/or MIPI SPI and charge pump clock frequencies.

Fig. 9. FDD FEM low band to high band desense in smartphones due to third harmonic.

To avoid any desense due to third harmonic, at least 90 dB isolation is required; this requires FEMiD shielding and special routing on the smartphone board.

Power amplifier

The RF PA is one of the core components for the FEM due to efficiency and linearity requirements, especially with the transition to 5G. With the adoption of 256 QAM for 5G UL the PA linearity expressed through EVM is started to be challenging to be met (<3.5%), although 5G allows 4–6 dB maximum power reduction (MPR) for 64/256 QAM 5G cyclic prefix–orthogonal frequency division (CP-OFDM) signals. Several techniques have been used to meet the efficiency and linearity requirements, the most extensively used and researched being Doherty and ET PAs [Reference Grebennikov17–Reference Kahn19]. The adoption of both techniques has been possible by advances in digital signal processing and technology scaling such as 7/3 nm FinFET. Doherty techniques provide high efficiency but have limitation in terms of broadband operation, operation in back-off mode, and load mismatch [Reference Zenteno, Isaksson and Händel20].

In mobile applications, such as smartphones, ET is used with a broadband PA with class E output match for low (LB), middle (MB), high (HB), and ultra-high (UHB) bands. The class E PA (Fig. 10) with parallel components has been introduced in [Reference Grebennikov and Jaeger21]. The optimum series feed inductance L and parallel capacitance C [Reference Grebennikov17, Reference Grebennikov and Jaeger21] can be obtained by the formulas:

(5)$$L = 0.732\displaystyle{R \over \omega }, \;\quad C = \displaystyle{{0.685} \over {\omega R}}, \;$$

Fig. 10. Power amplifier class E output match.

Lo and Co provide the series resonance, and for different bands, the structure might have 2Fo and 3Fo traps. As presented in [Reference Kim, Son, Kim, Jang, Oh and Park22, Reference Dinc, Kalia, Akhtar, Haroun, Cook and Sankaran23] this structure can be used with a differential transformer/balun. In [Reference Balteanu, Drogi, Hardik, Choi, Khesbak and Agarwal24] a HPUE PA with an ET has been presented. Using the same approach, a high resistivity CMOS 0.18 μm PA with a push–pull class E output stage and integrated fast error amplifier is presented in Fig. 11.

Fig. 11. LTE 4G/5G push–pull power amplifier.

The PA operates in class E [Reference Grebennikov and Jaeger21] and can use two or more push–pull structures to provide HPUE power capability (26 dBm) at the antenna for high 5G bands and operates with good efficiency at lower RF power levels. At lower power levels, few PA cores can be turned off. The ET structure is using an AC combiner which doesn't require a DC tracking loop [Reference Grebennikov17]. This is well suited for high modulation BW where the large AC coupling capacitor and the DC tracking loop will slow down the ET operation. One of the issues using ET [Reference Lie, Tsay, Hall, Nukala and Lopez14, Reference Grebennikov17] is the noise generated in the adjacent Tx channels especially for multiple Tx transmitters operating at the same time. To accommodate different Tx power levels and multiple Tx operating simultaneously, the ET can be operated in two modes:

1. (1) For lower modulation BW with less LTE RBs, the ET is done through the PA current bias (Itrck) as presented in Fig. 11 and the error amplifier is turned off.

2. (2) For higher RBs and multiple PA operating at the same time, the ET fast amplifier is turned on. Based on the signal modulation BW, different low frequency filters are selected on the Vdc_trck ports. This helps with the RF noise due to low frequency up conversion noise.

3. (3) Both Itrck and ET fast amplifiers are activated; assuming the envelope delays are well controlled.

Bias control for polar loop transmitters was introduced and described in [Reference Sowlati, Rozenblit, Pullela, Damgaard, McCarthy, Koh, Ripley, Balteanu and Gheorghe25] and can be extended to ET systems. For high modulation BW, such as in 3G/4G/5G, the polar approach is limited by the delay alignment for AM and PM paths. ET system is a hybrid approach between polar and linear modulation, where the delay alignment is critical but not as critical such as in polar systems. Assuming the delay is calibrated and manageable, Fig. 12 presents the static PA and ET operation under different bias (ET and PA gate bias) and therefore load line (Rload) conditions. Under ET operation the PA operates in back off mode, in class AB, B and A based on instantaneous RF power. When Itrck current is used and controlled, based on the instantaneous RF power the dynamic operating point is changing from load line 1 to load line 2. The operations in zones 3 and 4 are controlled by the Vdc_trck 1 and 2 voltages applied by the ET system based on a look-up table [Reference Grebennikov17]. Itrck can be generated using a similar look-up table, but this requires modem/transceiver resources. For the practical implementation Itrk is generated from Vdc_trck.

Fig. 12. Power amplifier and ET operation.

For HPUE, which is a requirement for mid-high 5G bands, two identical DC voltages (Vdc_dc1 and Vdc_dc2) are provided to the ET combiners which are located on the same FEMiD as presented in Fig. 13. This structure allows to achieve higher power at the PA output and accommodate the losses due to RF switches and acoustic filters. These losses are higher for 5G due to multiple antennas routing and higher linearity requirements for SOI RF switches.

Fig. 13. 5G power amplifier and ET.

The PA bias (Fig. 14) needs to provide the required current for the PA with low impedance to cancel the second-order input intercept point (IIP2) terms which might be upconverted and degrades the EVM and the ACPR/ACLR. The low impedance bias mitigates the current imbalances which produce ACPR/EVM degradation.

Fig. 14. Differential PA bias structure.

Also, the feedback bias amplifier (Fig. 15) must provide enough BW for the bias current. Due to these considerations a simple current feedback structure with source follower is used. To decrease the output impedance for quite large BW, a low Rs impedance is required but this increases the current consumption. From this perspective a programmable Rs is used.

Fig. 15. Differential PA bias structure detail.

The error amplifier which is part of ET is a current feedback amplifier with two stages and is presented in Fig. 16. The second stage is biased from Vdd_MLS where the voltage is set based on the modulation PAPR and the PA power level. This is determined with a comparator which senses the error amplifier output current through Isense pin. The Vctrl pin is used to compensate the current offset with a low-frequency feedback loop [Reference Drogi, Balteanu and Pehlke26].

Fig. 16. Current feedback error amplifier schematic.

When the PA is operated in the ET mode the voltage supply applied to the PA Vdd = Vdc_trck is following the envelope signal (instantaneous power level). The peak voltage applied is determined by the maximum power which must be delivered under ET for different peak to average power ratio (PAPR) waveforms [Reference Grebennikov17]:

(6)$${\rm PAPR} = 20\;\log \left({{\textstyle{{Vdc\_trck\_peak} \over {Vdc\_trck\_rms}}}} \right).$$

With the adoption of 5G there is an increase to 10.5 dB of the PAPR for all the waveforms. This determines the requirements for increased slew rate for the error amplifier and this is the main reason for using a current feedback amplifier.

To maintain a constant current and avoid any peaking and instability the error amplifier uses two current mode servo amplifiers Srv1 and Srv2 in the final stage as presented in Fig. 17. Using a current mode structure, the servo amplifiers provide a high operation BW and low propagation delay.

Fig. 17. Servo amplifier detail.

RF FEM SOI switches for 5G

The RF switches are integrated as separate dies to increase isolation and are fabricated in 0.18 μm SOI technology. The SOI switch structure is presented in Fig. 18.

Fig. 18. FEM RF SOI switch schematic.

For 5G FEMs, due to power increase for HPUE and increased PAPR for 5G NR waveforms, the number of series in FETs must be increased and determined by the formula:

(7)$$P_{\max } = \displaystyle{{V_{Tx\,\max }^2 } \over {2Z_0}} = \displaystyle{{2{( nV_{DS\_peak}) }^2} \over {Z_0}}, \;$$

where V_{DS_peak} is the peak NFET drain source voltage. For off state the peak RF voltage across FET drain-source for each transistor assuming equal voltage division is:

(8)$$\vert {V_{DS\_peak}} \vert = 2( V_{th}-V_{neg}) , \;$$

where V_th is the threshold FET voltage.

As presented in [Reference Zhu, Klimashov, Roy, Blin, Whitefield and Bartle27] the maximum breakdown voltage is determined by:

(9)$$BV_{\max } = \left({\sqrt {\displaystyle{{C_{ds}} \over {C_{gnd}}}} + 0.5} \right)BV_{FET}.$$

The insertion loss (IL) and the input intercept point (IIP3) [Reference Zhu, Klimashov, Roy, Blin, Whitefield and Bartle27, Reference Balteanu, Zhu and DiCarlo28] for a series-shunt switch are defined by the equations:

(10)$${\rm IIP}3 = 10\;\log \left({\displaystyle{{I_{sat}^2 {( R_{on} + 2Z_o) }^4} \over {4R_{on}Z_o^2 }}} \right) + 30, \;$$

(11)$${\rm IL} = 10\;\log \left[{{\left({1 + \displaystyle{{R_{on}} \over {2Z_o}}} \right)}^2 + \mathop {\left({\displaystyle{{2\pi C_{off}( R_{on} + Z_o) } \over 2}} \right)}\nolimits^2 } \right], \;$$

where R_on is the on-state channel resistance of the switch and V_pos is the positive voltage

(12)$$R_{on} = n\displaystyle{1 \over {\mu C_{ox}\displaystyle{w \over l}( {V_{\,pos}-V_{th}} ) }}.$$

Unfortunately for 5G the RF path losses are higher and the power requirements at the antenna are increased and therefore the switch performance requirements are more challenging. With two or more RF transmitters operating at the same time there are higher linearity requirements for antenna SOI switches as well as antenna tuning elements. For example, assuming sensitivity of a typical LTE 5 MHz (25RBs) as −101.5 dBm and 4.5 dB margin, the linearity requirement IIP3 to avoid jamming assuming Tx1 and Tx2 (Fig. 19) is given by:

(13)$$\eqalign{{\rm IIP}3 & = \displaystyle{{P_{Tx2} + 2P_{Tx1}-P_{IMD}} \over 2} = \displaystyle{{26 + 2 \times 23-( {-}106) } \over 2} \cr & = 89\;{\rm dBm}.}$$

Fig. 19. Cellular intermodulation distortions and cellular/WiFi desense.

Tx2 assumes 26 dBm which is required by the 3GPP with release #17 [10].

Assuming an external blocker at −30 dBm and a Tx2 UL signal at 23 dBm the switch linearity IIP2 is given by:

(14)$$\eqalign{{\rm IIP}2 & = P_{blk} + P_{Tx2}-P_{IMD} = {-}30 + 26-( {-}106) \cr & = { + } 102\;{\rm dBm}.}$$

To speed up the RF SOI switch transition, for a time response less than 1 μs (as required in 5G for low latency), a seven-stage ring oscillator with an activated clock doubler (for RF switching) is used. In order to generate V_pos and V_neg for the switch control [Reference Balteanu, Zhu and DiCarlo28], a negative/positive charge pump is used and presented in Fig. 20. This schematic has a ring oscillator structure with seven stages. For normal operation, the charge pump clock frequency is reduced using several dividers. Low clock frequency reduces the clock feedthrough into the RF PA in the Tx chain as well as Rx LNAs. For Tx applications there is a limit of the clock frequency reduction, due to effective reduction for V_neg and therefore a reduction for maximum RF power handling as expressed by equations (7) and (8).

Fig. 20. RF SOI switch positive/negative voltage generation with clock doubler.

RF 4G/5G acoustic filters

The transition from 3G/4G to 5G, with requirements for different RF transmitters’ coexistence, drives the use of several acoustic filters. For lower frequencies, up to 2 GHz, the filter requirements have been handled using SAW filters (Fig. 21). With the evolution to 5G and more HB and UHB FEMs, there is a high demand for BAW filters (Fig. 22). SAW filters can be used up to 2 GHz and must be temperature compensated (TC-SAW) to operate across a wide temperature range and to accommodate reduced guard bands, which is the case for 5G LB FDD filters.

Fig. 21. SAW acoustic filter structure.

Fig. 22. BAW acoustic filter structure.

BAW acoustic filters present quality factors of 2000 and 3000, which are higher compared with SAW filters, but require more precise steps in fabrication and are more costly. In Fig. 23, the measured TC-SAW acoustic filter characteristic for band n12 with 51–61 dB Tx rejection in Rx band is presented.

Fig. 23. TC-SAW acoustic filter response.

With the adoption of 5G for FDD LBs and the use of high modulation BWs, such as 64/256 QAM, the Tx and Rx filter flatness is crucial. There are many efforts in this area to reduce the spurious resonances in band caused by traverse modes [Reference Liu, Liu, Wang and Lam29], which produce different group delays based on the frequency of operation in Tx/Rx bands. The traverse mode suppression improves the ET delay calibration precision and reduces the time of ET calibration due to reduced number of frequency points for calibration.

Implementations and measurements

FEMs for smartphones end up into very high-volume products and therefore all the costs and hardware integration tests associated with the functionality must be considered. One of the RF Tx aspects is the calibration. Calibration also requires frequency equalization due to Tx and duplexer response for different modulation BW from 1.4 to 100 MHz. An ET calibration method where a baseband envelope signal is aligned using a triangular low intermediate frequency (IF)-modulated RF signal has been presented in [Reference Balteanu, Zhu and DiCarlo28] and used in current 4G/5G mobile devices. The delay is adjusted until the baseband-received signals have the same peak values. These RF peaks require an observation receiver to be processed. The observation receiver is a prerequisite of actual 4G/5G transceivers but the time for testing and calibration is increased. This contribution introduces a new calibration method where instead of the observation receiver a fast peak/rms detector is used, as presented in Fig. 24.

Fig. 24. FEM calibration for RF power and ET delay.

The peak detectors are very precise for PA power measurements (±0.2 dB) and the response BW can be adjusted from 1 GHz to lower frequencies such as 2 MHz. This allows for power measurements considering different modulation BWs and can be extended to mmWave FR2 measurements. Also, by using a baseband envelope signal with several peaks this can be aligned after detection and the ET delay calibrated. For higher modulation BW the ACLR is strongly depended on the delay mismatch as presented in Fig. 25, where the PA is set at lower power levels to avoid ACLR degradation due to heavy non-linearities.

Fig. 25. ACLR versus ET delay mismatch.

As general rule, for 20 MHz modulation BW the delay should be calibrated below ±0.5 ns to have good efficiency and at least −38 dBc ACLR for high power levels.

For one implementation, the PA, ET amplifier, and ET control and error amplifiers have been implemented using 0.18 μm CMOS with high resistivity technology with through silicon vias. Together with RF SOI switches and RF acoustic filters, all are integrated on the same substrate in a multi-chip-module. The worldwide transition of smartphones to 5G and the worldwide adoption for billions of people are one of main driving engine behind semiconductor industry. From this perspective the costs associated with the smartphone hardware as well as fabrication and field calibration must be considered. One of the important aspects for 5G Tx and Rx integration is the RF noise and noise measurements in production. The ET filters (Fig. 11) as well as charge pump noise mitigation have been implemented to provide enough margin from the design phase for the FEM noise in Rx band. In this way there is no need to check thoroughly during FEM production. RF noise mitigation is especially challenging for low frequency 5G bands where the duplex space is 30 MHz as in FDD bands n12 (699 MHz–716 MHz-UL/729 MHz–746 MHz-DL) and n13 (777 MHz–787 MHz-UL/746 MHz–756 MHz-DL). The Rx noise figure (NF) will degrade with 0.5 dB for a TX noise level lower than −120 dBm/Hz in Rx band. The FDD acoustic filter has a significant influence on the Rx band noise performance as the typical duplexer isolation between Tx and Rx is between 51 and 61 dB. Assuming the noise floor to be approximatively −174 dBm/Hz, for just 0.1 dB NF degradation, the noise after duplexer needs to be below −184 dBm/Hz. The noise measurements are shown in Fig. 26. For band n12 the measurements show low noise of −124 dBm/Hz at 30 MHz duplex space using an ET low-frequency filter with L_f = 39 nH and C_f = 1 nF (Fig. 10); P_out = 26 dBm. With the new TDD 5G bands [Reference Pehlke3] in the so-called C-band (n77, n78, and n79), there will be more BW available for high data rate but still the low-frequency bands will be used as 4G/5G LTE communication anchors. For example, right now the majority of 5G deployments use primarily 3.5 GHz TDD bands using NSA mode but will transition to SA. Table 2 presents the measurements for LTE 5 MHz, 710 MHz, Band n12 (30 MHz LTE BW), which exhibit very good efficiency for ACLR1 = −38 dBc, without digital predistortion, for 64 QAM. These measurements are done at the PA output. For 5G LBs, the power required at the antenna is 23 dBm and, with 5 dB MPR for 64 QAM CP-OFDM, the PA can be used in a LB FEM.

Fig. 26. Rx noise measurements for low-frequency 5G FDD bands – n13 and n12.

Table 2. CMOS 0.18 μm and GaAs PA and ET measurement results for LTE 10 MHz, 64 QAM, band n12

Similar implementation has been done using a GaAs PA (Fig. 27) with an integrated passive device (IPD) balun. In this case the ET bias and control circuitry is on a separate 0.18 μm CMOS die. The efficiency observed was better for GaAs versus CMOS, with almost 3–4% for the same ACLR and noise levels. For both implementations there is a penalty in terms of size using an IPD balun. At LBs (400–1200 MHz), using a balun helps with the second harmonic suppression for coexistence with other 5G transmitters and mitigates the low breakdown voltage for CMOS devices [Reference Balteanu, Drogi, Hardik, Choi, Khesbak and Agarwal24, Reference Aoki, Kee, Magoon, Aparicio, Bohn, Zachan, Hatcher, McClymont and Hajimiri30]. For size constrained applications, which is the case for 5G smartphones, the use of IPD baluns becomes practical for higher FR1 frequencies, such as n41, n77, n78, and n79 or FR2 mmWave frequencies.

Fig. 27. Power amplifier GaAs die for low bands 600–1200 MHz.

In Fig. 28, the measurements for a DP6T SOI switch used in the FEM are shown. The IIP3 is lower than 89 dBm as derived in equation (14). In this case the RFFE must operate under MPR or be used with different antenna route.

Fig. 28. DP6T insertion loss and IIP3 measurements for 50 Ohm load.