This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination. IEEE JOURNAL OF SOLID-STATE CIRCUITS 1 A Distributed Low-Noise Amplifier for Broadband Linearization of a Silicon Photonic Mach–Zehnder Modulator Navid Hosseinzadeh , Member, IEEE, Aditya Jain , Roger Helkey, Senior Member, IEEE, and James F. Buckwalter , Senior Member, IEEE Abstract— RF photonic components suffer from poor spurious-free dynamic range (SFDR) due to the high noise and distortion generated by a silicon photonic (SiP) Mach–Zehnder modulator (MZM). This work demonstrates a distributed silicon– germanium (SiGe) heterojunction bipolar transistor (HBT)-based low-noise amplifier (LNA) co-designed for linearization with a SiP MZM for a broadband radio-over-fiber (RoF) link. The SiGe LNA incorporates a distributed, tunable predistortion scheme that is inherently wideband and improves the thirdorder intercept point over an 18-GHz range. The assembled SiGe LNA and SiP MZM prototype demonstrates an SFDR as high as 120 dB · Hz2/3 at 9 GHz, a 14-dB improvement over previous SiP RF components or RoF links. Index Terms— Analog predistortion, distributed amplifier, driver, linearization, Mach–Zehnder modulator (MZM), microwave photonics, RF photonics, silicon–germanium (SiGe), silicon photonics (SiP), spurious-free dynamic range (SFDR), third-order intermodulation (IM3). Fig. 1. Top: block diagram of an RoF link with a SiGe LNA driving an MZM. Bottom: recently published SFDR results for SiP MZMs in RoF links. I. I NTRODUCTION R ADIO-OVER-FIBER (RoF) has the potential to support microwave and millimeter-wave communication with remote antenna heads. Remote antennas can improve wireless network reliability in environments with poor signal coverage; typically in situations with high RF path loss or blockage. In addition, RoF links are capable of capturing signals over a large instantaneous frequency range and transporting signals with low loss over optical fiber. RoF links are typically characterized by spurious-free dynamic range (SFDR), defined by (2/3) (IIP3-MDS), where IIP3 is the third-order input intercept point and MDS is the minimum detectable signal assuming 1-Hz bandwidth (BW). To support wideband RF Manuscript received May 5, 2020; revised August 30, 2020; accepted November 5, 2020. This article was approved by Guest Editor Daniel Friedman. This material is based on research conducted under AIM Photonics and sponsored by Air Force Research Laboratory via agreement number FA8650-15-2-5220. This article was presented at the IEEE Radio Frequency Integrated Circuit Symposium (RFIC 2019), Boston, MA, USA, June 2–7, 2019. (Corresponding author: Navid Hosseinzadeh.) Navid Hosseinzadeh and James F. Buckwalter are with the Department of Electrical and Computer Engineering, and Institute for Energy Efficiency, University of California at Santa Barbara, Santa Barbara, CA 93106 USA (e-mail: navid@ece.ucsb.edu; buckwalter@ece.ucsb.edu). Aditya Jain was with the Institute for Energy Efficiency, University of California at Santa Barbara, Santa Barbara, CA, USA. He is now with Seagate Technology, Edina, MN, USA. Roger Helkey was with the Institute for Energy Efficiency, University of California at Santa Barbara, Santa Barbara, CA, USA. He is now with Evidation Health, Santa Barbara, CA, USA. Color versions of one or more figures in this article are available at https://doi.org/10.1109/JSSC.2020.3038448. Digital Object Identifier 10.1109/JSSC.2020.3038448 and microwave communications applications, where complex modulation schemes are employed, the RoF link requires an SFDR exceeding 120 dB · Hz2/3 , compatible with RF receivers [1], [2]. Fig. 1 shows a conventional RoF link consisting of a lownoise amplifier (LNA) followed by an electro-optical (E/O) conversion using a Mach–Zehnder modulator (MZM). The RF-modulated optical signal is transported over fiber with little loss and converted back to the electrical domain through a photodetector. The linearity of the RoF link—as characterized by IIP3—is typically limited by Vπ , the voltage required to change the optical carrier phase by 180◦, of the MZM. While high Vπ improves IIP3, it reduces the RF link gain and increases the noise figure (NF) resulting in no overall SFDR improvement of the RoF link [3]. Optoelectronic materials offer different fundamental limits on the Vπ L product of the MZM, where L is the length of the MZM arms. For instance, InP MZMs typically exhibit 4 V·mm, whereas SiP MZMs exhibit 15 V·mm due to the weaker change in the index of refraction induced by plasma dispersion. Larger Vπ L of the SiP modulator requires long arms (L) to increase gain and reduce NF at the expense of frequency-dependent loss. Prior work illustrates that the SiP MZM has an SFDR of around 110 dB · Hz2/3 at 1 GHz that reduces at higher frequency [4], as shown in Fig. 1(b). Therefore, a 20-dB SFDR gap exists between current SiP MZMs [5]–[27] and the 120-dB · Hz2/3 target, particularly in microwave bands extending to 18 GHz. 0018-9200 © 2020 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See https://www.ieee.org/publications/rights/index.html for more information. Authorized licensed use limited to: Nvidia Corp. Downloaded on March 25,2021 at 03:50:26 UTC from IEEE Xplore. Restrictions apply. This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination. 2 IEEE JOURNAL OF SOLID-STATE CIRCUITS Fig. 2. Loaded and unloaded transmission line and corresponding lumped model for a small section. While SiP MZMs have higher Vπ L, Si processes offer integration of the E/O RoF receiver in a monolithic platform for larger scale RF systems. Consequently, this work attempts to overcome the physical IIP3 limits of the broadband MZM with an electronic linearization scheme. Prior optical linearization techniques include optical-based approaches, such as feedforward modulators and ring-assisted MZMs. While these approaches have demonstrated a 112-dB · Hz2/3 SFDR at 1 GHz and 94-dB · Hz2/3 SFDR at 20 GHz for a LiNbO3 MZM [22], they typically also require discrete electronic components [28]–[30]. Prior work [31]–[33] proposed predistorter drivers as a radio frequency integrated circuit (RFIC) for linearization of commercial LiNbO3 and electroabsorption modulators and reported up to 19.5-dB third-order intermodulation (IM3) distortion suppression. These improvements were limited to a maximum operation frequency of 1 GHz and did not explore linearization of SiP modulators, which require significant compensation to rival III–V optical modulators. Here, we present a broadband linearization approach based on electronic predistortion using the second-order intermodulation (IM2) products injection within a distributed amplifier to compensate for the nonlinearity of the SiP MZM and improve the SFDR to compete with LiNbO3 or InP MZMs. Three innovations are presented in this work. First, a distributed high-gain LNA is presented that covers 0.5–20 GHz to drive the MZM. Second, a distributed IM2 injection is demonstrated to reduce IM3 and linearize the cascade of the silicon–germanium (SiGe) LNA and MZM. Third, a co-design procedure for a segmented MZM (S-MZM) is described that predicts the broadband IIP3 and NF and supports the broadband linearization of the MZM. Prior work presented the circuit and initial measurements of the RoF link [34]. This article extends this earlier presentation with analysis of the linearity of the co-integrated SiP MZM and SiGe LNA to calculate the IIP3 of the RF photonic receiver. We also present new circuit analysis that is corroborated with measurement to describe the power consumption of the linearization approach given the IM3 reduction. A record 120-dB · Hz2/3 SFDR at 9 GHz and a response above 110 dB·Hz2/3 –18 GHz is demonstrated. The measured SiGe/SiP RoF indicates that the IIP3 improvement exists over more than 18 GHz, demonstrating the potential for broadband linearization. II. S-MZM S FOR B ROADBAND RF R ESPONSE SiP MZMs produce optical response through a change in the charge concentration in a p-n or p-i-n junction along the Fig. 3. IL of a loaded transmission line as a function of frequency for different diode resistances. optical waveguide. The MZM absorbs the reverse-biased depletion capacitor into a transmission line that can generally be modeled as a lumped circuit network where the LC ladder network predicts microwave propagation characteristics. In order to model loss, series resistance and shunt conductance are added to the LC network, where the former represents conductor loss and the latter represents substrate loss. Fig. 2 shows the unloaded and loaded transmission lines in the MZM and the corresponding lumped model unit cell, considering both conductor and dielectric loss. A small section of transmission line (z) is incorporated into the lumped model in Fig. 2, where z is the length of unit section. The attenuation constant per unit length and phase constant per unit length is C L 1 R +G (1a) αRF ∼ = 2 L C and √ βRF ∼ = ω LC. (1b) In an unloaded transmission line, the term Gz represents a large resistor, which models the conductor loss. The RF phase constant should be designed to match the optical travelingwave group velocity in a traveling-wave MZM (TW-MZM), whereas the attenuation constant should be minimized to reduce the losses along the length of the line. When the transmission line is loaded with a reverse-biased diode, we can use the RC network model in Fig. 2 to find the new value for G z. The series resistor R j can be converted into a shunt resistor using series-to-shunt conversion for RC networks. Consequently, we can find (1/G ) = R j (1+ Q s 2 ) = R j (1+(1/(ω2 C j 2 R j 2 ))), where Q s is the quality factor of the series RC circuit. Here, R j is defined as the total resistance that also includes the contact resistance and wiring (vias) resistance. We can see that G is a function of frequency, R j , and C j . With the new value of G z, Fig. 3 plots the attenuation constant of a 1-mm long transmission line as a function of frequency for different values of R j . Even a very small series junction resistance, αRF induces a limited BW and higher NF. Here, for simplicity, we assumed that R j is Authorized licensed use limited to: Nvidia Corp. Downloaded on March 25,2021 at 03:50:26 UTC from IEEE Xplore. Restrictions apply. This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination. HOSSEINZADEH et al.: DISTRIBUTED LNA FOR BROADBAND LINEARIZATION 3 where C J 0 is the zero-bias junction capacitance, Vbi is the built-in potential, and m is the grading coefficient of the junction. C0 and C1 are polynomial coefficients for (2). The capacitance nonlinearity is characterized by IIP2VQ and IIP3VQ and has been reported in prior work [35]. B. β Nonlinearity Fig. 4. Comparison of E/O integration strategies based on MZMs using traveling-wave design (top) and segmented design (bottom). independent of C j . This assumption is partially valid if the wiring resistance is dominant. However, in general, doping levels in the junction control the value of both R j and C j , where increasing doping levels will make R j smaller and C j larger. While smaller R j is preferred, larger C j means obtaining simultaneous 50- matching and E/O velocity matching is more difficult. Also, loss of the optical waveguides increases with higher doping levels. To improve the RoF link gain, a low Vπ demands long MZM arms. However, the p-i-n junction loss suggests that each arm should be under 1 mm to allow for a frequency response of 18 GHz. Therefore, the MZM transmission line of length L is broken into segments of less than 1 mm each. Consequently, the LNA should be distributed to allow a single transmission line input where the LNA buffers the signal as well as produces a differential drive signal to the two arms of the MZM. The LNA is a buffer between the transmission line and reverse bias diodes. The insertion loss (IL) is improved, resulting in lower NF for the system at high frequency. A four-segment MZM shown in Fig. 4 is designed in a 65-nm SiP technology, where each segment includes a differential p-i-n optical phase shifter to confine the optical field in the waveguide. While the SMZM improves the SFDR at high frequencies compared to a TW-MZM, it does not improve the SFDR beyond 110 dB · Hz2/3 [35]. Increasing the SFDR further demands an active linearization approach to improve IIP3 that does not penalize NF. III. D ISTORTION A NALYSIS OF S I P E/O I NTERFACES The optical carrier travels with propagation constant γopt = αopt + jβopt , where βRF ∼ = βopt . To modulate the optical carrier with the RF signal, the propagation constant has a linear dependence on the applied voltage such that αopt = a0 + a1 Vi and βopt = b0 + b1 Vi , where b1 = (1/Vπ L). We will examine how the SiP MZM deviates from this model and produces distortion in the following. The change in the charge produced by the depletion capacitance produces a change in the phase shift through the waveguide that is not necessarily linear but rather modeled by a higher order polynomial. Rather than considering the voltage dependence on the phase constant, a charge relationship is used here to isolate the RF and optical sources of distortion β(Q) = m 2π 2π n eff ≈ qk Q k λ λ (3) k=1 where qk is polynomial coefficient for the nonlinear index of refraction and λ is the light wavelength. The phase nonlinearity is characterized by IIP2 Qβ = |q1 /q2 | and IIP3 Qβ = (|4q1/3q3 |)1/2 . C. Vπ Nonlinearity The optical transmission is determined by the differential phase shift between the two MZM arms. In addition, a dc voltage is applied to the electrode that results in a static phase shift αDC = 2α0 + α1 (V A + V B ) [35] Pout e−αDC L (cosh(2α(Q)L) + cos(2β(Q)L − φDC )) = Pin 2 (4) m k where α(Q) = k=1 αk Q is defined as the polynomial modeling α as a function of charge. The generalized model presented here characterizes the MZM by IIP2β P and IIP3β P to incorporate the transfer function from the junction capacitance nonlinearity and phase nonlinearity of p-n-junction phase shifters [35], [36]. If (4) is assumed to be ideal with linear β with respect to the applied voltage and quadrature bias point, distortion is caused by the sin(x) transfer of Vi to Pout conversion. The output of optical transmission is modeled with a third-order memoryless nonlinearity Pout π 1 π 3 3 3 −αDC L = c1 v i − c3 v i = e vi − v i . (5) Pin Vπ 6 Vπ Consequently, the√intermodulation distortion is characterized from IIP3V π = 2 2(Vπ /π). For a 3-mm-long SiP, the equivalent IIP3 is 26 dBm. Therefore, if we are dominated by the sine nonlinearity, the polynomial model is a memoryless, weakly nonlinear function and linearization of this characteristic is the primary role of broadband predistortion circuitry. A. C–V Nonlinearity Charge in the waveguide is modulated through a reversebiased depletion capacitance, Q = C(V )V , where CJ0 (2) C(V ) = (1 − V /Vbi )m D. Cascaded Nonlinear Analysis The block diagram in Fig. 5 represents both the amplitude and phase distortion in the MZM. Neglecting the amplitude path, which primarily generates only even-order distortion Authorized licensed use limited to: Nvidia Corp. Downloaded on March 25,2021 at 03:50:26 UTC from IEEE Xplore. Restrictions apply. This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination. 4 IEEE JOURNAL OF SOLID-STATE CIRCUITS Fig. 5. Block diagram representing a nonlinear model for distortion in the SiP MZM. Fig. 7. IIP3 and gain of the RC network using Volterra series for different configurations, where L = Nseg × L seg , as a function of frequency. If we also include the diode depletion capacitance, the short length more significantly reduces the OIP3, particularly as the frequency increases. While we conclude that OIP3 improves for longer length, higher frequency loss also increases as discussed in Section II, which results in a tradeoff between linearity and BW. IV. L INEARIZATION T ECHNIQUES FOR S I P C OMPONENTS Fig. 6. OIP3 as a function of MZM length for various cases considering different sources of nonlinearity, found from cascade analysis. when biased at quadrature and driven differentially, a distortion cascade analysis for IIP3 combines each source of nonlinearity 1 IIP32MZM 2 |A1 |2 C02 2π |A1 |2 C02 |A1 |2 λ n1 = + + 2 2 IIP3VQ IIP3Qβ IIP3β P 2 2π 3 C1 |A2 | C02 |A1 |2 λ n 2 +C1 |A2 | + + 2 IIP2 Qβ IIP2β P 2π λ n1 (6) where A1 ( j ω) and A2 ( j ω) are Volterra kernels found for (2) and presented in the Appendix. The cascade analysis indicates how the components interact to produce a frequencydependent distortion due to the capacitance of the waveguide. Considering that the RF gain of the link is G RF = sin φDC e−αDC L Pin Ro ∼ Pin Ro = Vπ Vπ (7) where is the responsivity in A/W of the photodetector and Ro is the output resistance. The OIP3 is calculated from the product of IIP3MZM and G RF and is plotted in Fig. 6. When considering only the Vπ nonlinearity, the OIP3 is independent of length since IIP3 increases with Vπ , but the gain reduces with Vπ . When adding the βopt nonlinearity, the OIP3 reduces for shorter MZM arms since a much larger voltage must be produced across the MZM diode capacitance and the large voltage produces a more nonlinear response in the phase constant. We propose three distinct RF linearization techniques for the O/E and E/O components specific to each source of nonlinearity presented in Section III. A. C–V Linearization To mitigate the nonlinearity due to C–V variation in the depletion capacitance, the segments of the MZM allow the RC time constant of the segment to be tailored to reduce the generation of distortion. First, the design incorporates more fixed capacitance at each segment, which reduces the contribution of the nonlinear C–V component. In addition, a smaller resistance can be used to drive each segment to reduce the large-voltage variation that results in intermodulation distortion. In addition, biasing the junctions at higher reverse bias voltages will further reduce the capacitance and improves linearity. In a TW-MZM, junction capacitance and loading on the electrodes are optimized for E/O velocity matching and matching the characteristic impedance to the source impedance. These constraints move to other parts of the design, the driver design, and hence, phase shifter loading can be optimized for linearity. IIP3 due to C–V variation can be characterized by [35] IIP3CV = 1 + 5ω2 Ri2 C02 + 4ω4 Ri4 C04 4 2 4 3ω Ri C2 C2 1 + ω2 Ri2 C12 + 2C0 (8) where Ri is the load resistance. Fig. 7 shows IIP3 and gain due to capacitance nonlinearity found from (8) and compared cases with one 4-mm segment with two 2-mm and four 1-mm segments. As expected, gain roll-off increases as we Authorized licensed use limited to: Nvidia Corp. Downloaded on March 25,2021 at 03:50:26 UTC from IEEE Xplore. Restrictions apply. This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination. HOSSEINZADEH et al.: DISTRIBUTED LNA FOR BROADBAND LINEARIZATION Fig. 8. 5 Concept of MZM linearization by third-order predistortion. increase the segment length and, hence, capacitance. The IIP3 improvement is due to fixed load resistance being used for all three cases, which keeps voltage swings the same as well. Increasing a number of segments furthermore and making them shorter will lead to improved BW. However, at the same time, power consumption of the driver chip is increased proportional to a number of segments. Hence, we choose to have a minimum number of segments for a specific BW. Here, we targeted >20-GHz BW, and, based on simulation results in Fig. 7, we chose four segments. B. Kerr Effect Linearization Prior work has demonstrated the linearization of the optical phase shift with induced voltage through engineering of the second-order optical response with respect to the electric field induced in the p-n junction of the waveguide [37]. A p-i-n junction diode is formed in the waveguide, which introduces a Kerr effect (χ-2) dependence on the optical phase. Balancing the plasma dispersion and Kerr effect modulation linearizes the phase constant with applied voltage. Nonetheless, SiP DC Kerr effect MZMs have reported only slight improvement of the SFDR at low frequency and no improvement higher frequency bands [4], [38], [39]. This is presumably due to the linearity being limited by other sources of nonlinearity, such as MZM transfer function and nonlinear junction capacitor. C. Vπ Linearization Since the arcsin(x) transfer function is frequency independent, a broadband predistortion circuit could linearize the MZM. The system shown in Fig. 8 is a cubic predistorter for linearization of the MZM. Here, a second path generates a cubic term electronically, which is then weighted by w and added back to the main signal before applying to the MZM. The predistorted signal is v x = v i + wv i3 = v i · 1 + wv i2 . (9) Applying this predistortion signal to the modulator results in Pout = c1 v i + (wc1 − c3 )v i3 − c3 wv i5 − c3 w2 v i7 − c3 w3 v i9 . Pin (10) The third-order term is eliminated when w = (c3 /c1 ) and the cubic polynomial generates a number of additional higher order terms. The effectiveness of this cancellation is demonstrated in the OIP3 plot as a function of MZM length for different values for w, represented in Fig. 9. For each device Fig. 9. OIP3 as a function of MZM length when the system is limited by different sources of nonlinearity (dashed line) and after applying predistortion for different third-order component weights (solid line). Each solid line is plotted for a different weight of injected IM3 component. length, an optimum w produces OIP3 that exceeds the original OIP3 (21 dBm) realized from the sine nonlinearity. Predistortion results in peaking for particular device lengths. Adding other sources of nonlinearity discussed in Sections III-A and B results in red and blue curves in Fig. 9. Comparing these two figures indicates how secondary sources of nonlinearity, charge and phase shift, affect the overall OIP3. Designing a broadband cubic term generator is challenging since the cubic nonlinearity generates signal components over three times the original signal BW. Another approach to IM3 generation is to mix IM2 with the original signal as indicated in the final equality in (9) to imitate a cubic term generator response, which we refer to as IM2 injection [40]. A broadband IM2 generator is realized from the intrinsic nonlinearity of an heterojunction bipolar transistor (HBT) device as is typically used in a frequency doubler. Since the squaring of an RF signal at frequency ωRF generates components at dc and 2ωRF , we eliminate the components at 2ωRF with a low-pass filter. The remaining enveloped is modulated with the original RF signal and weighted to produce the desired IM3 term. Fig. 10 shows a conceptual block diagram of the proposed distributed, analog predistortion LNA. The first block is the LNA with high gain to set the NF of the RoF and relax constraints on the length of MZM. The signal is then split between two paths. In one path, the RF signal is amplified, while through an auxiliary path, the RF signal is squared and filtered to generate a relatively low-frequency IM2 component. Next, the IM2 component is mixed with the original RF signal to generate an IM3 component that can be weighted and added to the original RF signal to cancel the IM3 components generated by the MZM. By adjusting the IM3 weight (w), the RoF can reach an IIP3 higher than the inherent MZM IIP3, as shown in Fig. 9. V. C IRCUIT I MPLEMENTATION The distributed predistortion stage shown in Fig. 10 is realized in an HBT technology. Fig. 11 shows the schematic Authorized licensed use limited to: Nvidia Corp. Downloaded on March 25,2021 at 03:50:26 UTC from IEEE Xplore. Restrictions apply. This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination. 6 IEEE JOURNAL OF SOLID-STATE CIRCUITS Fig. 10. Conceptual illustration of the broadband linearization scheme with block diagram of distributed IM3 injection. Fig. 12. Circuit schematic illustrating the highly linear differential driver and IM2 generator block. filter the IM2 products. The weighting and mixing function is produced by mirroring the IM2 current into the bias current of the high-voltage differential driver. The output current of the IM2 generator block, which is based on a frequency doubler, is composed of a dc and IM2 component I B v id 2 (11) IIM2 = I B + 2 VT 2 Fig. 11. Circuit schematic illustrating the specific components including the single-ended LNA, wideband active balun, variable delay transmission line. of the LNA, active wideband balun, and tunable delay input transmission line. The wideband LNA is a resistively degenerated cascode stage with inductive peaking to extend the BW. Resistive degeneration extends the broadband frequency response [41]. Next, an active balun produces differential RF signals with less than 1-dB amplitude imbalance and under 5◦ of phase imbalance up to 25 GHz [35]. To match the optical traveling-wave velocity in the MZM with the RF traveling wave, the transmission line delay is adjusted over a small range with a varactor. The distributed design allows delay control between the RF input traveling-wave line and the IM2 generation between stages. Fig. 12 shows the schematic of the linear, high-voltage differential cascode driver, which amplifies the RF signal to 2 Vpp to drive the p-i-n phase shifter segments in the S-MZM along with the IM2 injection circuitry. The IM2 injection circuitry taps the differential signals and uses an inductively degenerated frequency doubler to square and where I B is the dc bias current, VT is the thermal voltage, and v id is the amplitude of differential input voltage. If we drive the IM2 current generated with two closely spaced tones at ω1 and ω2 , the current terms are produced at ω1 − ω2 and ω1 + ω2 . Both IM2 components can be injected into the driver to produce IM3 components. Filtering of the sum components occurs due to the frequency response of the IM2 generator circuit. In Fig. 13, the inset plot shows that the difference frequency current is constant to 18 GHz. The sum frequency is roughly 8 dB lower in the broadband IM2 generator. The IM2 current is plotted in Fig. 13 for tones located at approximately 2 and 20 GHz. The tone spacing is 100 MHz for the simulation, and the dc bias current is 10 mA. The maximum IM2 current under this bias is 4 mA. The output voltage of a differential BJT amplifier as a function of differential input voltage follows the tanh(x) function. Therefore, predistorted output voltage for the circuit in Fig. 12 is 2 I B v id v id . (12) tanh Vout = R L IEE + I B + 2 4 2V VT T Representing tanh(x) as a linear Taylor series equivalent 2 v id I B v id 1 v id 3 Vout = R L IEE + I B + 2 − 2VT 24 VT VT 4 Authorized licensed use limited to: Nvidia Corp. Downloaded on March 25,2021 at 03:50:26 UTC from IEEE Xplore. Restrictions apply. (13) This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination. HOSSEINZADEH et al.: DISTRIBUTED LNA FOR BROADBAND LINEARIZATION 7 Fig. 13. Current at ω2 − ω1 and ω2 + ω1 for the IM2 generator as a function of input power, at 2 and 20 GHz. Fig. 14. OIP3 of the driver as a function of frequency for different IM2 injection currents. where we ignored higher order terms generated by the tanh function. The desired predistorted signal is v id VSW v id 3 Vout = VSW + IB RL − (14) 2VT 3 2VT heater, and photodiode to capture the nature of the ideal and nonlinear behavior. With accurate behavioral modeling of the SiP MZM, the improvement in the OIP3 can be characterized. Fig. 15 shows the co-simulation test bench, including the driver and S-MZM. A four-segment optical model for the SiP MZM is constructed in Cadence and is shown in Fig. 15. The sources of distortion are presented in Sections III-A–III-C. The nonlinear characteristics for the capacitance nonlinearity and phase and amplitude nonlinearity are included for the p-i-n phase shifters. The capacitance change per applied reverse bias voltage is illustrated to the bottom left and was measured for the SiP MZM with an LC R meter. The capacitance changes from 0.7 to 1.14 pF/mm. In addition, the change in the index of refraction and attenuation constant with charge stored in the optical waveguide is plotted on the left of Fig. 15. To establish the linearity in terms of an RF output power, the MZM is connected to an ideal photodiode with the responsivity of 1 A/W to produce an output current as terminated by a 50-. The circuit schematic of the four-stage SiGe distributed IM2-injection LNA is shown in Fig. 15 based on the predistortion technique from Fig. 10. Each driver stage is co-designed, specifically for the SiP S-MZM stages. The input transmission line distributes the RF signal to the four LNA stages. The complete circuit schematic includes the LNA, active balun, driver, and IM3 generation. We simulate the OIP3 of the RoF link in Fig. 16 to indicate the benefit of the feedforward, distributed IM3 injection. The OIP3 is calculated in terms of RF power at the output of the photodiode found from E/O co-simulation of driver and MZM. In the absence of any predistortion, the OIP3 is around 5 dBm and rolls off around 10 GHz. The use of predistortion with injection of IM2 components into the same stage indicates that the OIP3 is improved by 10 dB. However, the OIP3 is limited to under a few GHz due to the timing mismatch between the RF signal and the IM3 products. By injecting the IM2 component to the consecutive distributed stage, the time delay can be adjusted to achieve where VSW = R L (IEE + I B ). Note that I B is the desired weighting control that allows selection of the expansive IM3 products and can compensate for distortion generated by the driver or the following MZM. Generally, we assume IEE + I B ≈ IEE . Examination of tanh indicates that it also produces undesirable third-order distortion. Consequently, by calculating gain and IIP3 = ((4/3)|(a1/a3 )|)1/2 from (13), the OIP3 for the driver stage as a function of IM2 injection dc current is IEE + I B . (15) OIP3 = 2VSW IEE − 2I B The OIP3 is tuned through I B and can cancel the selfgenerated distortion for I B = IEE /2. The injected IM2 current is relatively large suggesting high power consumption for the IM2 generator. For instance, a 1-V maximum swing requires 15 mA of I B current. An alternative approach is to linearize the output driver such that the OIP3 can be improved to compensate for the MZM nonlinearity. To include the role of R E (IEE + I B )R E 3 1+ OIP3 OIP3 . (16) VT Fig. 14 shows OIP3 of the driver as a function of frequency for different IM2 injection current amplitudes demonstrating broadband improvement of OIP3. The average simulated OIP3 of the driver is about 43 dBm, an increase of more than 14 dB in the absence of IM2 injection. VI. C O -S IMULATION OF E LECTRICAL AND O PTICAL IC S To co-simulate the RFIC and PIC, a Verilog-A library was developed for optical components, including laser, silicon waveguide, splitter and combiner, p-i-n-junction phase shifter, Authorized licensed use limited to: Nvidia Corp. Downloaded on March 25,2021 at 03:50:26 UTC from IEEE Xplore. Restrictions apply. This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination. 8 Fig. 15. IEEE JOURNAL OF SOLID-STATE CIRCUITS Electrooptic co-simulation test-bench based on Verilog-A models for SiP and BiCMOS process models. Fig. 16. OIP2 and OIP3 found from electrooptic co-simulation of driver and MZM for three different cases: 1) IM2 off; 2) IM2 on and injected to the same stage; and 3) IM2 on and injected to the next stage. Fig. 17. Simulated OIP3 variation after employing IM3 injection: with temperature (top) and with NPN corner parameter (bottom). a broadband linearization. The OIP3 now remains above 15 dBm to 18 GHz. Consequently, the SFDR should be improved by more than 6 dB using the distributed predistortion scheme. The co-simulated OIP2 is also plotted in Fig. 16. While there is a reduction in the OIP2 due to the injection of the IM2 term, all three cases indicate that the minimum OIP2 remains greater than 30 dBm from 0.5 to 20 GHz. Stability of the linearization technique to temperature variation as well as process variation is investigated. Fig. 17 (top) shows the effect of temperature on the simulated OIP3, where the temperature is varied from 0 ◦ C to 100 ◦ C with 25 ◦ C steps. The result shows good thermal stability with low sensitivity to temperature change. Moreover, Fig. 17 (bottom) exhibits the effect of NPN transistors corner parameter variation on the linearized OIP3. Increasing this parameter will result in higher speed transistors with higher currents. The result here shows low sensitivity up to 12 GHz. In the future, a closed-loop system can be designed to control the bias of the IM2 generator block based on the performance for real-time optimization and calibration, which improves the stability furthermore. While a one-time calibration for each module would work well, a calibration that is running in the background can be employed to monitor the power of the input signal and possible blocker and adjust the bias of the IM2 generator blocks and other stages according to the power levels and frequency. Such a system will help to lower the power consumption and maximize the performance at the same time. VII. M EASUREMENTS R ESULTS The SiP S-MZM is fabricated in the 65-nm AIM silicon photonic (SiP) process and the broadband, low-noise driver is fabricated in the GlobalFoundries 130-nm SiGe BiCMOS technology. The SiP S-MZM area is 5.5 × 0.65 mm2 and driver chip area is 3.95 × 1.38 mm2 and is shown in Fig. 18. The SiGe RFIC was designed using RF output GSSG pads with matching pitch to the MZM. The PCB assembly is shown in Fig. 18 with the SiGe distributed LNA and SiP MZM chips Authorized licensed use limited to: Nvidia Corp. Downloaded on March 25,2021 at 03:50:26 UTC from IEEE Xplore. Restrictions apply. This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination. HOSSEINZADEH et al.: DISTRIBUTED LNA FOR BROADBAND LINEARIZATION 9 Fig. 18. Measurement setups for link characterization with QAM signal (top left) and linearity measurements (top right), micrograph of the assembly (bottom left), SiP S-MZM (middle), and LNA chip fabricated in 130-nm SiGe BiCMOS technology (bottom right). along with two 1550-nm lensed fibers used for optical interface with edge couplers. The two chips are assembled side-byside on a PCB to minimize chip-to-chip wirebond length. The power consumption of the driver is 1.70 W (425 mW per stage) without IM3 predistortion and increases by 60 mW with the use of IM3 predistortion to provide the 10 to 20-dB IIP3 improvement. Most of the power consumption is consumed in the segment driver, which is designed to produce 2 Vpp across the load resistor using a 3.3-V supply. Fig. 18 shows a block diagram of the measurement setup with PNA-X, photodetector, micro-positioners, microscope for fiber alignment, and the PCB. Two different setups are used: one for small-signal and large-signal characterizations using PNA-X and one for link characterization with modulated signals using an arbitrary waveform generator and oscilloscope. A 1550-nm low-RIN (−168 dB/Hz) laser drives the MZM in all measurements followed by a manual polarization controller. The measured loss for optical edge couplers is about 3 dB and the total optical loss of MZM is less than 10 dB, including light coupling loss. The output of the MZM is sent to a 50-GHz dual-band 1310-/1550-nm photodetector, which generates an RF signal to the PNA-X. Fig. 19 plots the measured S-parameters for the RoF link with an external high-linearity photodetector. The link gain is plotted against the measurement of a similar SiP TW-MZM in the same technology with the same length. The active driver improves the gain by 26 dB at 3 GHz and improves the gain by 40 dB above 15 GHz. This supports the advantage of the segmented design to improve the high-frequency response. The simulated and measured input return losses are also shown in Fig. 19(middle), and the measured S11 is better than 10 dB to more than 20 GHz. The measured group delay is plotted in Fig. 19(bottom) and indicates a relatively low variation up to 20 GHz, with a maximum of 0.24 ns at 16–18 GHz. Fig. 20 shows the measured output power spectrum for two tones at 1.9 and 2 GHz with and without IM2 injection. The two tones are generated by the PNA-X and combined into the driver. After applying IM2 injection and tuning the Fig. 19. Measured RoF link gain compared to a single TW-MZM (top), simulated and measured input return loss (middle), and group delay as a function of frequency (bottom). Fig. 20. Measured power spectrum for two tones at 1.9 and 2 GHz at input of the driver. IM3 reduction manually in the driver, the IM3 components decrease by more than 23 dB, whereas an IM2 term at 3.9 GHz increases by 8 dB. In addition, the HD3 components are reduced by 10 dB, which demonstrates that the IM3 predistortion operates over a wide BW. Authorized licensed use limited to: Nvidia Corp. Downloaded on March 25,2021 at 03:50:26 UTC from IEEE Xplore. Restrictions apply. This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination. 10 IEEE JOURNAL OF SOLID-STATE CIRCUITS Fig. 22. Measured IIP2 versus frequency before and after IM2 injection (top) and measured IIP3 versus frequency with and without IM3 injection at two bias points (bottom). Fig. 21. Input power sweep at 5 and 10 GHz with and without IM3 injection for fundamental and IM3 components. The IM3 measurement is also characterized by the fundamental and IM3 over a range of input powers. Fig. 21 shows the input power sweep at 5 and 10 GHz and indicates the predistortion improves the IIP3 by 10 and 14 dB, respectively. The weight of the IM3 component is controlled by the bias current of the IM2 generator and a slight increase in the injected fundamental power is observed. Fig. 22 plots the IIP3 with and without IM2 injection at two different bias currents to demonstrate the linearization BW. Over the entire 1–20-GHz range, the IIP3 is increased relative to the LNA without IM2 injection by at most 17 dB. The bias on the IM2 injection is adjusted to achieve higher IIP3 improvement above and below 10 GHz. The bias current of the IM2 generator block was 3.4 and 5.6 mA to get maximum IIP3 improvement in the 1–10- and 10–18-GHz bands, respectively. For broadband (larger than an octave) systems, IIP2 is also a relevant metric. Since the IM2 generator injects IM2 components, the IIP2 is plotted as a function of frequency with and without IM2 injection. Notably, IIP2 does not substantially change over the frequency range with the IM2 bias current. Fig. 23 plots the measured NF compared to the estimated NF of a Si-based TW-MZM. The LNA improves the NF by more than 20 dB with a minimum 13.6 dB at 3 GHz. From the measured IIP3 and NF, the SFDR is 120 dB · Hz2/3 at 9 GHz, a 19-dB improvement over earlier work and 22-dB improvement at 17 GHz. A3 ( j ωa , j ωb , j ωc ) = Fig. 23. Measured NF of the RoF link compared to the estimated NF of a Si-based TW-MZM (top) and calculated SFDR from measurement data and comparison with [22] (bottom). The RoF link is also tested with 2-Gbd QPSK, 16-QAM, and 64-QAM modulated signals at a carrier frequency of 11 GHz. For QPSK at 4 Gb/s, 16-QAM at 8 Gb/s, and 64-QAM at 12 Gb/s, the measured EVMs are 2.6%, 3.7%, and 6.5%, respectively, which are measured without employing any digital predistortion. This work is compared to earlier reported linearized MZM components or RoF links in Table I. The presented work demonstrates the best SFDR for a SiP RF photonic link over the highest frequency range and puts silicon MZMs at a similar j (ωa + ωb + ωc )Ri (C1 A12 + C32 A1 ( j ωa )A1 ( j ωb )A1 ( j ωc )) 1 + j (ωa + ωb + ωc )C Ri Authorized licensed use limited to: Nvidia Corp. Downloaded on March 25,2021 at 03:50:26 UTC from IEEE Xplore. Restrictions apply. (17c) This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination. HOSSEINZADEH et al.: DISTRIBUTED LNA FOR BROADBAND LINEARIZATION 11 TABLE I S TATE - OF - THE -A RT H IGH SFDR RF P HOTONIC M ODULATORS ACKNOWLEDGMENT Fig. 24. Measured QPSK (top), 16-QAM (middle), and 64-QAM constellations, all at 2 GS/s. This material is based on research conducted under AIM Photonics. The authors would like to thank Erman Timurdogan and Zhan Su of Analog Photonics for providing the p-i-n phase shifter building block and GlobalFoundries for access to the 8XP process technology. They would also like to thank Prof. Schow for helpful discussions and Prof. Bowers and Prof. Rodwell for access to measurement equipment. They acknowledge Integrand for access to EMX software. R EFERENCES capability with reported LiNbO3 modulators. In addition, this is the first reported work based on an integrated electronicphotonic solution for RoF with small size and weight. VIII. C ONCLUSION This work reports a co-designed SiGe LNA and SiP-based MZM for SFDR of up to 120 dB · Hz2/3 and operating up to 20 GHz. A broadband analog predistortion scheme is introduced that forwards IM2 distortion products into the SiGe driver to eliminate IM3 distortion generated by the SiP MZM. The results demonstrate that SiP MZMs are capable of meeting high-performance criteria for RoF links and meet or surpass commercial LiNbO3 and III–V modulator performance. Prospects for future co-integration of SiP and RF electronics on a single platform could support the linearization of photonic devices. A PPENDIX The Volterra kernels are defined by 1 (17a) 1 + j ωa C Ri −1 j C1 Ri (ωa + ωb )A1 ( j ωa )A1 ( j ωb ) A2 ( j ωa , j ωb ) = 2 1 + j (ωa + ωb )C0 Ri (17b) A1 ( j ωa ) = and, (17c), as shown at the bottom of the previous page, where A12 = 1 (A1 ( j ωa )A2 ( j ωb , j ωc ) 3 + A1 ( j ωb )A2 ( j ωa , j ωc ) + A1 ( j ωc )A2 ( j ωa , j ωb )). (17d) [1] R. W. Ridgway, C. L. Dohrman, and J. A. Conway, “Microwave photonics programs at DARPA,” J. Lightw. Technol., vol. 32, no. 20, pp. 3428–3439, Oct. 15, 2014. [2] N. Hosseinzadeh, A. Jain, R. Helkey, and J. F. Buckwalter, “RF silicon photonics for wideband, high dynamic range microwave and millimeter-wave signal processing,” in Proc. IEEE 18th Topical Meeting Silicon Monolithic Integr. Circuits RF Syst. (SiRF), Jan. 2018, pp. 41–44. [3] V. J. Urick, K. J. Williams, and J. D. McKinney, Fundamentals of Microwave Photonics. Hoboken, NJ, USA: Wiley, 2015. [4] C. G. Bottenfield, V. A. Thomas, and S. E. Ralph, “Silicon photonic modulator linearity and optimization for microwave photonic links,” IEEE J. Sel. Topics Quantum Electron., vol. 25, no. 5, pp. 1–10, Sep. 2019. [5] V. J. Urick, M. S. Rogge, P. F. Knapp, L. Swingen, and F. Bucholtz, “Wide-band predistortion linearization for externally modulated longhaul analog fiber-optic links,” IEEE Trans. Microwave Theory Techn., vol. 54, no. 4, pp. 1458–1463, Jun. 2006. [6] C. Zhang, P. A. Morton, J. B. Khurgin, J. D. Peters, and J. E. Bowers, “Ultralinear heterogeneously integrated ring-assisted Mach–Zehnder interferometer modulator on silicon,” Optica, vol. 3, no. 12, pp. 1483–1488, Dec. 2016. [7] J. Ding, S. Shao, L. Zhang, X. Fu, and L. Yang, “Method to improve the linearity of the silicon Mach-Zehnder optical modulator by doping control,” Opt. Express, vol. 24, no. 21, pp. 24641–24648, Oct. 2016. [8] C. Zhang, P. A. Morton, J. B. Khurgin, J. D. Peters, and J. E. Bowers, “Highly linear heterogeneous-integrated Mach-Zehnder interferometer modulators on Si,” Opt. Express, vol. 24, no. 17, pp. 19040–19047, Aug. 2016. [9] J. Cardenas et al., “Linearized silicon modulator based on a ring assisted Mach Zehnder inteferometer,” Opt. Express, vol. 21, no. 19, pp. 22549–22557, 2013. [10] C. Xiong et al., “A linear push-pull silicon optical modulator,” in Proc. Frontiers Opt. Washington, DC, USA: Optical Society America, 2014, p. FM3A.4. [11] Y. Zhou et al., “Linearity measurement of a silicon singledrive push-pull Mach-Zehnder modulator,” in Proc. CLEO, 2015, p. SW3N.6. [12] A. Ayazi, T. Baehr-Jones, Y. Liu, A. E.-J. Lim, and M. Hochberg, “Linearity of silicon ring modulators for analog optical links,” Opt. Express, vol. 20, no. 12, pp. 13115–13122, Jun. 2012. Authorized licensed use limited to: Nvidia Corp. Downloaded on March 25,2021 at 03:50:26 UTC from IEEE Xplore. Restrictions apply. This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination. 12 IEEE JOURNAL OF SOLID-STATE CIRCUITS [13] A. M. Gutierrez et al., “Silicon-based electro-optic modulators for linear and nonlinear radio-over-fiber applications,” in Proc. IEEE Int. Topical Meeting Microw. Photon., Sep. 2012, pp. 168–171. [14] A. M. Gutierrez et al., “High linear ring-assisted MZI electrooptic silicon modulators suitable for radio-over-fiber applications,” in Proc. 9th Int. Conf. Group IV Photon. (GFP), Aug. 2012, pp. 57–59. [15] L. Chen, J. Chen, J. Nagy, and R. M. Reano, “Highly linear ring modulator from hybrid silicon and lithium niobate,” Opt. Express, vol. 23, no. 10, pp. 13255–13264, May 2015. [16] M. Streshinsky et al., “Highly linear silicon traveling wave MachZehnder carrier depletion modulator based on differential drive,” Opt. Express, vol. 21, no. 3, pp. 3818–3825, Feb. 2013. [17] D. Zhu, J. Chen, and S. Pan, “Multi-octave linearized analog photonic link based on a polarization-multiplexing dual-parallel MachZehnder modulator,” Opt. Express, vol. 24, no. 10, pp. 11009–11016, May 2016. [18] X. Han, X. Chen, and J. Yao, “Simultaneous even- and thirdorder distortion suppression in a microwave photonic link based on orthogonal polarization modulation, balanced detection, and optical sideband filtering,” Opt. Express, vol. 24, no. 13, pp. 14812–14827, Jun. 2016. [19] S. Li, X. Zheng, H. Zhang, and B. Zhou, “Highly linear radioover-fiber system incorporating a single-drive dual-parallel Mach– Zehnder modulator,” IEEE Photon. Technol. Lett., vol. 22, no. 24, pp. 1775–1777, Dec. 15, 2010. [20] X. J. Meng and A. Karim, “Microwave photonic link with carrier suppression for increased dynamic range,” Fiber Integr. Opt., vol. 25, no. 3, pp. 161–174, Jul. 2006. [21] Z. Chen et al., “SFDR enhancement in analog photonic links by simultaneous compensation for dispersion and nonlinearity,” Opt. Express, vol. 21, no. 18, pp. 20999–21009, Sep. 2013. [22] R. Zhu, X. Zhang, D. Shen, and Y. Zhang, “Ultra broadband predistortion circuit for radio-over-fiber transmission systems,” J. Lightw. Technol., vol. 34, no. 22, pp. 5137–5145, Nov. 15, 2016. [23] S. Jin, L. Xu, and Y. Li, “Quantum well Mach-Zehnder intensity modulator with enhanced linearity,” in IEEE MTT-S Int. Microw. Symp. Dig., May 2016, pp. 1–3. [24] B. Liu, J. E. Bowers, Y.-J. Chiu, A. Keating, J. Piprek, and J. Shim, “Analog characterization of low-voltage MQW traveling-wave electroabsorption modulators,” J. Lightw. Technol., vol. 21, no. 12, pp. 3011–3019, Dec. 2003. [25] A. Rao et al., “High-performance and linear thin-film lithium niobate Mach-Zehnder modulators on silicon up to 50 GHz,” Opt. Lett., vol. 41, no. 24, pp. 5700–5703, Dec. 2016. [26] P. A. Morton, J. Cardenas, M. Lipson, and J. B. Khurgin, “Miniature, linearized silicon photonics modulators for phased array systems,” in Proc. IEEE Int. Topical Meeting Microw. Photon. (MWP), Oct. 2013, pp. 275–277. [27] N. Hosseinzadeh, A. Jain, K. Ning, R. Helkey, and J. F. Buckwalter, “A 1 to 20 GHz silicon-germanium low-noise distributed driver for RF silicon photonic Mach-Zehnder modulators,” in IEEE MTT-S Int. Microw. Symp. Dig., Jun. 2019, pp. 774–777. [28] W. Jiang et al., “A linearization analog photonic link with high thirdorder intermodulation distortion suppression based on dual-parallel Mach–Zehnder modulator,” IEEE Photon. J., vol. 7, no. 3, pp. 1–8, Jun. 2015. [29] A. Karim and J. Devenport, “High dynamic range microwave photonic links for RF signal transport and RF-IF conversion,” J. Lightw. Technol., vol. 26, no. 15, pp. 2718–2724, Aug. 1, 2008. [30] A. Nikolov, D. Guenther, W. Liu, R. Cendejas, and R. Dutt, “Advancements and challenges for photonic components for avionic interconnects,” in Proc. IEEE Avionics, Fiber-Optics Photon. Technol. Conf. (AVFOP), Oct. 2013, pp. 5–6. [31] R. Sadhwani and B. Jalali, “Adaptive CMOS predistortion linearizer for fiber-optic links,” J. Lightw. Technol., vol. 21, no. 12, pp. 3180–3193, Dec. 2003. [32] Y. Chiu, B. Jalali, S. Garner, and W. Steier, “Broad-band electronic linearizer for externally modulated analog fiber-optic links,” IEEE Photon. Technol. Lett., vol. 11, no. 1, pp. 48–50, Jan. 1999. [33] J. Okyere, K. Yu, K. Entesari, and S. Palermo, “A fifth-order polynomial predistortion circuit for Mach-Zehnder modulator linearization in 65 nm CMOS,” in Proc. Texas Symp. Wireless Microw. Circuits Syst. (WMCS), Mar. 2017, pp. 1–4. [34] N. Hosseinzadeh, A. Jain, K. Ning, R. Helkey, and J. Buckwalter, “A 0.5-20 GHz RF silicon photonic receiver with 120 dB•Hz2/3 SFDR using broadband distributed IM3 injection linearization,” in Proc. IEEE Radio Freq. Integr. Circuits Symp. (RFIC), Jun. 2019, pp. 99–102. [35] N. Hosseinzadeh, A. Jain, K. Ning, R. Helkey, and J. F. Buckwalter, “A linear microwave electro-optic front end with SiGe distributed amplifiers and segmented silicon photonic Mach–Zehnder modulator,” IEEE Trans. Microwave Theory Techn., vol. 67, no. 12, pp. 5446–5458, Dec. 2019. [36] N. Hosseinzadeh, A. Jain, R. Helkey, and J. Buckwalter, “Sources of RF intermodulation distortion in silicon photonic modulators,” in Proc. IEEE Avionics Vehicle Fiber-Opt. Photon. Conf. (AVFOP), Nov. 2018, pp. 1–2. [37] E. Timurdogan, C. V. Poulton, M. Byrd, and M. Watts, “Electric fieldinduced second-order nonlinear optical effects in silicon waveguides,” Nature Photon., vol. 11, no. 3, p. 200, 2017. [38] C. G. Bottenfield, V. A. Thomas, and S. E. Ralph, “Microwave photonic links: Optimizing SiP modulator design and operation,” in Proc. Opt. Fiber Commun. Conf. (OFC), 2019, pp. 1–3. [39] A. Jain et al., “A high spur-free dynamic range silicon DC kerr ring modulator for RF applications,” J. Lightw. Technol., vol. 37, no. 13, pp. 3261–3272, Jul. 1, 2019. [40] S. Lou and H. C. Luong, “A linearization technique for RF receiver front-end using second-order-intermodulation injection,” IEEE J. Solid-State Circuits, vol. 43, no. 11, pp. 2404–2412, Nov. 2008. [41] K. Fang, C. S. Levy, and J. F. Buckwalter, “Supply-scaling for efficiency enhancement in distributed power amplifiers,” IEEE J. Solid-State Circuits, vol. 51, no. 9, pp. 1994–2005, Sep. 2016. Navid Hosseinzadeh (Member, IEEE) received the M.Sc. degree in electrical engineering from the Sharif University of Technology, Tehran, Iran, in 2014, and the Ph.D. degree in electrical and computer engineering from the Department of Electrical and Computer Engineering, University of California at Santa Barbara, Santa Barbara, CA, USA, in 2020. He is currently a Post-Doctoral Researcher with the University of California at Santa Barbara. His research interests include RF/millimeter-wave integrated circuit (IC) design in silicon and compound semiconductor technologies, silicon photonics integrated circuits, and quantum computers. In 2019, he was a Research and Development Intern with PsiQuantum Corporation, Palo Alto, CA, USA. Dr. Hosseinzadeh serves as a Reviewer for IEEE M ICROWAVE AND W IRELESS C OMPONENTS L ETTERS , IEEE T RANSACTION ON C IRCUITS AND S YSTEMS —I: R EGULAR PAPERS , IEEE T RANSACTION ON C IR CUITS AND S YSTEMS —II: E XPRESS B RIEFS , and IEEE T RANSACTION ON M ICROWAVE T HEORY AND T ECHNIQUES . Aditya Jain received the B.Eng. degree from the University of Birla Institute of Technology and Science, Pilani, India, in 2010, and the Ph.D. degree from the Department of Electrical Engineering, Iowa State University, Ames, IA, USA, in 2016. From 2016 to 2018, he was a Post-Doctoral Fellow with the University of California at Santa Barbara (UCSB), Santa Barbara, CA, USA. His research interests include silicon photonics and metamaterials. Authorized licensed use limited to: Nvidia Corp. Downloaded on March 25,2021 at 03:50:26 UTC from IEEE Xplore. Restrictions apply. This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination. HOSSEINZADEH et al.: DISTRIBUTED LNA FOR BROADBAND LINEARIZATION Roger Helkey (Senior Member, IEEE) received the B.S. degree in engineering from Caltech, Pasadena, CA, USA, and the Ph.D. degree in optoelectronics from the University of California, Santa Barbara, CA, USA. Previously, he was Associate Director with the UCSB Institute for Energy Efficiency, Santa Barbara. He is currently a Senior Data Engineer at Evidation Health, Santa Barbara. At Calient Networks, Goleta, CA, USA, he developed optical switches as a lowpower alternative to electrical switches for data center and telecommunications applications. He has authored four book chapters and received 42 patents. He is a Fellow of the OSA. 13 James F. Buckwalter (Senior Member, IEEE) received the Ph.D. degree in electrical engineering from the California Institute of Technology, Pasadena, CA, USA, in 2006. From 1999 to 2000, he was a Research Scientist with Telcordia Technologies, Morristown, NJ, USA. In 2004, he joined the IBM T. J. Watson Research Center, Yorktown Heights, NY, USA. In 2006, he joined the Faculty of the University of California at San Diego, La Jolla, CA, as an Assistant Professor, where he became an Associate Professor in 2012. He is currently a Professor of electrical and computer engineering with the University of California at Santa Barbara, Santa Barbara, CA. Dr. Buckwalter was a recipient of the 2004 IBM Ph.D. Fellowship, the 2007 Defense Advanced Research Projects Agency Young Faculty Award, the 2011 NSF CAREER Award, and the 2015 IEEE MTT-S Young Engineer Award. Authorized licensed use limited to: Nvidia Corp. Downloaded on March 25,2021 at 03:50:26 UTC from IEEE Xplore. Restrictions apply.