1. INTRODUCTION
Recent advances in photonic materials, devices, and architectures have shown considerable promise for applications in artificial intelligence (AI) and neuromorphic computing [1,2]. One of the major motivations for photonics in the context of AI is the ability to multiplex information using many optical wavelengths and process it simultaneously (using wavelength-division multiplexing) in an analog photonic processing unit at high speeds. This approach allows for parallel processing at high speeds, offering a viable alternative to traditional computing hardware. This is particularly beneficial for overcoming limitations such as energy consumption, computing density, and scalability that challenge current digital processors.
To maximize the potential of photonic computing, most photonic processing units are designed around a reconfigurable array of memory cells. These cells are programmed to implement a matrix of trained weights [3–7]. Matrix–vector multiplication is then achieved through parallel multiply-accumulate operations as a vector of optical inputs passes through the memory array and is read out by photodetectors at the output of the array [8]. While the optical input signals can be modulated at very high speeds with established electro-optic effects (free carrier, Kerr, Franz–Keldysh, etc.), the memory cells themselves are often much slower. This disparity between the memory cell and modulator is often due to a variety of competing requirements [9], such as low insertion loss, low static power consumption, small device footprint, high bit precision, and high dynamic range.
Fig. 1. Overview of multifunctional photonic memory concept. (a) Illustration and working mechanism of multifunctional photonic memory leveraging a PN junction microheater and modulator. Coarse nonvolatile tuning of the PCM is achieved by applying a forward-biased pulse to locally heat the waveguide while reverse biasing the PN junction modulates carriers in the depletion region for volatile fine-tuning. (b) Measured current density as a function of voltage applied to the waveguide-integrated PN junction. Biasing the PN junction below ∼1.5V enables volatile tuning without significant heating of the PCM. (c) Modeling comparison of heat generated in PIN, PN, and ?-doped waveguide-integrated microheaters. All simulations use a fixed +5V forward bias and 1 µm spacing between ?++/?++ regions. (d) Optical microscope image of a fabricated microring array (bottom, 25 µm scale bars) and magnified image of a single microring with integrated multifunctional memory cell (top left, 10 µm scale bar). Cross-sectional view of a PN microheater with PCM deposited in the oxide window (top right).
Furthermore, for practical scalability of the memory array, low static power consumption and small device footprint are critically important. This has spurred research into nonvolatile optical mechanisms for storing information on-chip. Among various demonstrated [9–11] nonvolatile optical effects—e.g., mechanical [12], memristive [13,14], ferroelectric [15], magneto-optic [16], etc.—phase-change materials (PCMs) have emerged as particularly promising [17]. PCMs offer a large refractive index change (Δ?∼0.5−2), compact footprint (1–10 µm), long-term retention (∼10years [18]), and multilevel storage capabilities (up to 7 bits per cell [19]). However, the major challenges facing these devices are (1) their slow electronic reconfigurability of 50 ns–100 µs; (2) power-hungry operation (10 nJ–1 µJ); and (3) limited endurance of 1000–10,000 cycles. It has been noted previously that photonic computing architectures leveraging phase-change memory can be significantly less power efficient and slower than their digital counterparts when programming energy and time are considered [20,21]. Adding additional tunability to the memory cell without changing the state of the PCM is one approach which could address several of these issues provided the tuning is fast and efficient. While prior works have explored using the volatile response in PCMs [22,23] or an underlying microheater [24] to add this tunable functionality, these approaches are limited by slow response times and high static power dissipation.
Here, we propose and demonstrate a multifunctional photonic memory cell that addresses these challenges by combining the benefits of nonvolatile phase-change materials with the efficient and high-speed tunability of foundry-processed modulators in a microring structure. Notably, this multifunctional cell utilizes a waveguide-integrated PN junction for both programming the PCM via localized heating under forward bias and high-speed fine-tuning of the optical weights under reverse bias conditions. Our co-integrated memory with both nonvolatile and volatile functionality aims to extend the domains of photonic computing beyond accelerating machine learning inference to fast and efficient in situ training and online learning.
2. RESULTS
Our device concept is shown in Fig. 1(a), where we have integrated two PN junctions into the left and right sides of a microring resonator (MRR) with independent electrical control. This independent control allows us to selectively switch the side of the MRR containing the PCM to improve efficiency and localize heat generation. When an applied electrical pulse exceeds the threshold voltage of the PN junction (∼0.7V), the PN junction is forward-biased and can locally heat the PCM, triggering a phase transition and enabling coarse-tuning of the optical resonance. Fine-tuning of the resonance can be achieved through reverse biasing the PN junction, which changes the carrier concentration (and, therefore, effective index) in the waveguide. While this free carrier effect is volatile and relatively small, combining this high-speed fine-tuning with the slow-speed coarse-tuning of PCMs allows for both precise and efficient control over the optical weights. This is particularly useful in the context of on-chip training, where small changes in the optical weights can be rapidly implemented through the reverse-biased PN junction. Only when the accumulated weight change exceeds the tuning range of the PN junction is the PCM reprogrammed, thus improving the memory cell endurance and the energy efficiency and speed of the training process.
In Fig. 1(b), we plot the measured current density of a fabricated PN junction in both the forward and reverse bias conditions. Removing carriers from the waveguide by changing the depletion width results in a red shift of the resonance since for a silicon waveguide, Δ?eff=−?Δ?−?Δ?0.8, where Δ?eff is the change in effective index of the waveguide, Δ? and Δ? are the changes in the electron and hole concentrations, and ? and ? are constants which depend on the wavelength [25]. It is worth noting that volatile tuning of the resonance can also be achieved through forward biasing the junction between 0 V and ∼1.5V, causing a blue shift due to carrier injection. However, for forward biasing beyond the threshold voltage, heating of the waveguide and free carrier absorption begins to dominate, which reduces the efficiency and extinction ratio. For large forward voltages (>4V), the heating can be significant enough to reach the crystallization and melting temperatures of the PCM, enabling reversible switching.
Figure 1(c) compares the simulated heat generation from three foundry-compatible microheater designs employing PIN, PN, and ?-doped microheaters. These simulations were performed using COMSOL Multiphysics using the same material properties and methods we have explored in prior work [26]. For all designs, the separation between the heavily doped ?++ and ?++ regions was fixed at 1 µm (i.e., 1 µm long intrinsic, PN, and n-doped regions), and +5V was applied in the forward-biased direction. The observed peaks in the heat generation are mainly due to discontinuities in the silicon thickness (i.e., waveguide versus slab height) and changes in doping concentration and type. For the PIN and PN junctions, we observe heat generation due to both carrier recombination and Joule heating, while for the n-doped microheater, only Joule heating occurs due to the presence of only one carrier type.
In terms of heating efficiency, the ?-doped heater maximizes the heat generated near the waveguide while minimizing heating in the heavily doped contacts compared to the PIN and PN microheaters. However, fine-tuning with an ?-doped microheater will be dominated by Joule heating, which limits fine-tuning to sub-MHz speeds. Compared to the PIN microheater, the PN microheater generates more heat at the same forward bias voltage and can also be reverse biased to enable fine-tuning of the optical weights. While PIN and n-doped microheaters have been successfully used in the past to electrically switch PCMs [27–29], volatile tuning is only possible in the forward-biased direction [24]. This type of fine-tuning requires a constant current to be applied, which increases the static power dissipation of the device to several mW levels and reduces efficiency in the system. Additionally, for PIN junctions, the maximum tuning speed is limited by carrier recombination in the intrinsic region, resulting in slow (∼15MHz) volatile tuning speeds [24]. Using our PN junction design, we can achieve fine-tuning of the optical weights with sub-nanosecond speeds and with less than 80 pA of leakage current at a reverse bias of −8V, resulting in less than 640 pW of power dissipation per device.
Our multifunctional memory cell is particularly well suited for the broadcast and weight architecture [30], which requires accurate tuning of the resonance position for each optical weight [31–33]. Four multifunctional photonic memory cells sharing common bus waveguides for both positive and negative weighting are shown in Fig. 1(d). These devices were fabricated at Advanced Micro Foundry (AMF) using their active silicon photonic process (AMFSiP). Oxide windows etched down to the waveguide by AMF were used to integrate Ge2Sb2Se4Te1 (GSST) and Sb2Se3 after device fabrication [visible in the magnified image of a single memory cell in Fig. 1(d)]. Sputtering was then used to deposit 30 nm of either GSST or Sb2Se3, followed by a 30 nm oxide cladding to prevent oxidation (see Section 4). A cross-sectional view of the final device is illustrated in Fig. 1(d).
To achieve reliable switching, all devices underwent a post-deposition anneal to ensure the PCM is initially in the crystalline state. The importance of this annealing step can be seen in Fig. 2. In Fig. 2(a), we show the crystallization process of a memory cell with 5 µm of GSST without an initial hot plate anneal. The PN microheater was used to crystallize the GSST using 10 ms pulses of increasing amplitude ranging from 4 V to 5 V. As the pulse amplitude increases above the glass transition temperature of the GSST, the resonance experiences a blue shift while the extinction ratio decreases. However, the resonance eventually shifts toward longer wavelengths, as expected when the GSST is fully crystallized at higher pulse amplitudes. This blue shift can be explained by a topographical change in the distribution of GSST where material migrates away from the waveguide when heated above the glass transition temperature, as illustrated in Fig. 2(a), thus lowering ?eff and shifting the resonance to shorter wavelengths. We confirm that this process is indeed topographical rather than ablation or oxidation by reversibly switching the device between the amorphous and crystalline states after initial crystallization [see Fig. 2(b)]. Amorphization was achieved using 100 µs pulses with an amplitude ranging from 7 V to 8 V. While the device can still function as nonvolatile memory, the reduced material above the waveguide after this reflow process lowers the maximum wavelength shift that can be achieved between the crystalline and amorphous states.
Fig. 2. Effects of annealing step on deposited PCM (Ge2Sb2Se1Te4). (a) Observed initial spectral blue shift prior to reaching the fully crystallized state due to reflow of the as-deposited PCM above the glass transition temperature. (b) After reflow, reversible switching is achieved, demonstrating a topographical change in the PCM rather than a structural or chemical change. (c) Observed spectral red shift of both resonances after annealing the entire chip on a hot plate. The large red shift is expected and indicates minimal reflow of the PCM after annealing. (d) Spectrum of the microring with reamorphized PCM matches well with the initial as-deposited state, demonstrating minimal topographical changes after hot plate annealing.
To address this issue, we first perform a 200°C anneal on a hot plate for 10 min. The spectra before and after this anneal step can be seen in Fig. 2(c) for another set of memory cells with the same geometry and length of GSST as in Figs. 2(a) and 2(b). When the GSST is crystallized, there is a clear shift to longer wavelengths and a reduced extinction ratio after annealing. This is expected for crystalline GSST since the refractive index and loss both increase relative to the as-deposited amorphous state. We observe minimal reflow in these devices, as evidenced by the ability to reamorphize the GSST to a resonance position and extinction ratio very close to the initial as-deposited state [Fig. 2(d)]. While these results shown are for GSST, we observed the same behavior in Sb2Se3 samples with and without the post-deposition anneal step.
To demonstrate nonvolatile programming, we used memory cells with 2 µm of PCM (GSST and Sb2Se3) and set the optical weights using a series of amorphization pulses. Figures 3(a) and 3(b) show the spectra of two memory cell arrays containing four MRRs, each before and after programming. In these experiments, Ring 1 was fully amorphized, Ring 4 was left in the crystalline state, and Rings 2 and 3 were set to an arbitrarily chosen partially crystalline state. While the resonance shift is slightly smaller for the devices with Sb2Se3, the lower loss results in a higher extinction ratio in the crystalline state [Fig. 3(b)] compared to the rings with the same length of GSST [Fig. 3(a)]. This tuning range can be further increased by increasing the PCM length. While the loss begins to dominate for longer GSST lengths (see Fig. 2), we observed reasonably high extinction ratios (>12dB) in devices with 5 µm long Sb2Se3, which were able to achieve spectral shifts >800pm (see Fig. S2 in Supplement 1).
Fig. 3. Nonvolatile tuning of photonic memory. (a),(b) Nonvolatile programming of arbitrary optical weights using memory cell arrays with four MRRs containing (a) GSST and (b) Sb2Se3. For these experiments, Ring 1 was fully amorphized, Rings 2 and 3 were partially amorphized, and Ring 4 was left in the crystalline state. (c) Incremental amorphization of GSST using a PN microheater (probe wavelength indicated by the dashed line). (d) Measured blue shift and (e) transmission state of Ring 1 from (c) as a function of amorphization pulse number. Amorphization pulses were 100 µs in width with an increasing voltage ranging from 5.6 V to 6.4 V.
Coarse control of the optical weights is shown in Fig. 3(c), where amorphization pulses of increasing amplitude are used (100 µs pulse width with amplitude ranging from 5.6 V to 6.4 V). Here, we achieve 14 distinct levels (3.8 bits) and shift the resonance position by 270 pm [see Fig. 3(d)]. This large spectral shift allows us to achieve an insertion loss of −1.3dB and an extinction ratio of 22 dB between the fully crystalline and amorphous states. In Fig. 3(e), we plot the transmission measured at a probe wavelength of 1512.05 nm [dashed line in Fig. 3(c)] on a linear scale.
Having demonstrated coarse tuning of the optical weights, we also show volatile fine-tuning by biasing PN junctions below the threshold voltage. Figure 4(a) shows the measured resonance shift of the PN junction (19.2 µm arc length) under reverse bias conditions. As the reverse bias voltage increases, carriers are removed from the waveguide, and the depletion region increases. This results in a red shift of the spectra. Our measurements agree well with the analytical model found in Chrostowski and Hochberg [25] for a PN modulator with ?- and ?-doping in the junction equal to 5×1017cm−3. As the PN junction in our current devices only occupied 42% of the ring, this modulation efficiency could be further improved by a factor of ∼2× simply by increasing the fraction of our memory cell covered by the PN junction.
Fig. 4. Volatile tuning of photonic weight. (a) Measured resonance shift as a function of reverse bias voltage. The red line shows the expected resonance shift based on the analytical model of carrier modulation in the PN depletion region. (b) Shift in MRR resonance as a function of forward bias voltage. After ∼1.5V, Joule heating begins to dominate and shifts the resonance back to longer wavelengths. (c) Static power consumption versus resonance shift for the PN junction in both forward and reverse bias configurations. Reverse biasing the device provides >10,000× better power efficiency. (d) Normalized frequency-dependent electro-optic response of the PN junction in both forward and reverse bias conditions. Carrier recombination limits the frequency response to ∼200MHz under forward bias, while the RF bandwidth of the measurement setup limits the response in reverse bias.
It is also possible to achieve volatile tuning using a forward bias voltage. In this configuration, carriers are injected into the waveguide (rather than removed) and cause a blue shift of the resonance due to plasma dispersion. The change in carrier density under a forward bias voltage can be much larger than under a reverse bias, and therefore, the resonance shift is greater as well. In Fig. 4(b), we plot the measured shift in resonance for the case of the 10 µm PN microheater under forward bias. After the applied bias exceeds ∼2× the threshold voltage of the PN junction, Joule heating begins to dominate, and the resonance shifts to longer wavelengths. This can be modeled using the following equation:
where ? is the applied voltage; ? is the current; and ?, ?, and ? are fitting parameters. The first term in Eq. (1) corresponds to the power dissipated by the PN junction (Joule heating), while the second term is dependent on the carrier density in the waveguide (plasma dispersion). After fitting to the measured wavelength shift in Fig. 4(b), we obtain ?=156.6pm/mW, ?=−89.4pm/mA, and ?=0.756. Since holes tend to dominate changes in refractive index at carrier concentrations below ∼1019cm−3, this value of ? is in good agreement with the commonly used value of 0.8 from Reed et al. [34]. The contributions due to Joule heating and plasma dispersion are plotted in Fig. 4(b) as red and green lines, respectively, with their combined effects shown in black.
While the volatile tuning range is greater for forward biasing compared to reverse biasing, the static power dissipation is much worse. In Fig. 4(c), we compare the measured power dissipation as a function of the achievable resonance shift. While the resonance shift is limited to tens of picometers for the reverse bias case, within this tuning range, the power required to maintain this volatile state is over 4 orders of magnitude less than the case for forward bias. This is important when considering future scaling to large memory arrays where the static power required to fine-tune optical weights could become a source of significant energy loss for the system.
Fig. 5. Multifunctional operation. (a) Demonstration of both nonvolatile and volatile tunability in our multifunctional memory cell. The PCM is cycled between partially amorphous and fully crystalline states with increasing amorphization pulse amplitude. Blue regions show volatile tuning of the PN junction before returning the PCM to the crystalline state. (b) Zoomed-in region from (a) showing both volatile and nonvolatile tunability. Small discontinuities indicated by vertical arrows are due to noise from electrical relays used to switch between instruments. (c) Nonvolatile cyclability of our memory cell between the fully amorphous and fully crystalline states. The resonance of the ring in the crystalline state blue shifts during cycling (see inset), indicating the endurance is limited by a degradation of the crystalline state.
In addition to power dissipation, the maximum modulation speed under forward bias is limited by carrier recombination to several nanoseconds for a PN junction [35] and tens of nanoseconds for a PIN junction [24]. This is again orders of magnitude worse than the reverse biased case where the modulation speed is limited to tens of GHz by the RC time constant of the PN junction [25]. We compare the frequency response of our PN junction in forward and reverse bias in Fig. 4(d). A bias tee was used to provide a DC offset to the RF signal coming from a vector network analyzer with 7.5 GHz bandwidth. For the case of a positive DC offset voltage (forward biased condition), we see that the ?3dB point is ∼200MHz, which agrees well with other forward-biased silicon PN modulators [35]. For the case of a negative DC offset, the roll-off occurs at 7.1 GHz since carriers are swept out of the PN junction under a reverse bias and carrier recombination is negligible. While this is already more than a 400× improvement compared to volatile tuning of PCM memory using a PIN junction [24], we expect the response of the memory cell to exceed 20 GHz under reverse bias based on the estimated RC time constant and photon lifetime in the ring [25]. However, in the case of Fig. 4(d), the RF bandwidth of the bias tee limits the maximum frequency response of our measurement setup.
In a final measurement, we demonstrate the multifunctionality of our memory cell by combining both nonvolatile coarse-tuning of the PCM with volatile fine-tuning using the reverse-biased PN junction. These measurements use a ring with 5 µm of Sb2Se3 integrated onto a 10 µm long PN microheater. In Fig. 5(a), we program the memory cell to a nonvolatile state using eight different amorphization pulses ranging from 3.63 V to 4.44 V in amplitude and 5 µs in duration (see Section 4). After each amorphization pulse, a reverse bias sweep is applied to the PN junction followed by a forward-biased crystallization pulse to return the PCM to its fully crystalline state (2.8 V with 5 ms duration). A zoomed-in view of one of these cycles can be seen in Fig. 5(b). Note, the spikes in transmission observed in Figs. 5(a) and 5(b) are due to the thermo-optic response of the MRR during the 5 ms crystallization pulse. We can see that the PN junction is able to function reliably both as a microheater for switching the PCM and as a phase modulator even after several switching cycles. Later during endurance measurements, we found no degradation in the PN junctions even after over 10,000 switching cycles (see Fig. S5 in Supplement 1). In Fig. 5(c), we show the cycling endurance of our memory cell when switching the Sb2Se3 between its fully amorphous and fully crystalline states. In these measurements, we perform a spectral measurement of the ring resonance after every 100 cycles [see inset of Fig. 5(c)] and realign the probe laser as needed. During these cycling measurements, we observe a significant spectral blue shift in the resonance of the ring when the Sb2Se3 is in the crystalline state. This indicates a reduction in the fraction of crystalline material over time as observed by others [36]. While improving the cycling endurance of these PCMs is still an active area of research [37,38], reducing the total number of rewrites by leveraging volatile tuning as we have demonstrated could be an effective method for prolonging the lifetime of these devices in addition to providing higher precision and greater flexibility to PCM-based photonic memories.
3. CONCLUSION
We have demonstrated a multifunctional photonic memory cell with an integrated PN microheater and modulator, functionalized with PCMs. This design allows both coarse-tuning via the PCM and fine-tuning by reverse biasing the PN junction to enhance the resolution, programming speed, and lifetime of the memory cell. Compared to prior work [24], we improve the volatile modulation speed and tuning efficiency by more than 400× and 10,000×, respectively, while introducing a new class of waveguide-integrated microheater with low optical loss and efficient heat generation. Notably, our memory cell is compatible with commercial photonics foundry process flow with the ability to incorporate PCMs using a simple back-end-of-line post-process step. These key innovations are important for the future development of computational photonic memory arrays, which are fast, efficient, and scalable.
4. METHODS
A. Device Fabrication
The devices were fabricated at AMF, Singapore, using an active silicon photonic process (AMFSiP) with oxide windows etched down to the waveguides for back-end-of-the-line materials integration. The MRR employs a partially etched waveguide geometry, consisting of a 130 nm rib section on top of a 90 nm slab, which optimizes both optical confinement and electrical conductivity. This design enables efficient optical mode propagation while facilitating carrier injection and depletion through the integrated PN junction. Finite-difference eigenmode simulations show that the fundamental TE mode is primarily confined within the rib section (please see Fig. S1 in Supplement 1), with an effective index of approximately 2.57 at 1550 nm. The partial etch depth was chosen based on the fabrication process offered by AMF. The 130 nm etch provides strong optical confinement and thus smaller bend radii, while the remaining 90 nm slab ensures efficient electrical connectivity to the PN junction for both forward and reverse bias operation. Further optimization could explore asymmetric waveguide geometries or alternative doping profiles to enhance modulation efficiency while maintaining low optical loss. 30 nm of Sb2Se3 or GSST was deposited from single targets using an AJA Orion-3 ultra-high vacuum sputtering system at room temperature. The patterning of the PCM cells was performed using electron beam lithography on an Elionix ELS-G100 system with Ma–N 2403 negative resist, followed by CF4 reactive-ion etching. 30 nm of SiO2 was subsequently deposited via sputtering to protect the PCM from oxidation.
B. Measurement Setup
A tunable fiber laser (Santec TSL-550) was utilized to generate input light with wavelengths ranging from 1500 to 1630 nm for the experiments. A 16-channel fiber array was employed to couple light into and out of the chip via on-chip grating couplers. A two-channel custom RF Probe (S-G-S) from GGB Industries with a bandwidth of up to 40 GHz was utilized for modulation of the PN junctions. To facilitate the capture and real-time recording of optical data from the chip, the output of a fiber-coupled photodetector (Newport 2011-FC) was connected to a BNC-2110 data acquisition board from National Instruments. To capture the spectral response of the rings, the wavelength of the tunable laser source was swept from 1500 to 1630 nm at a speed of 100 nm/s, and 100,000 samples were captured from the output photodetector by the DAQ board for the entire sweep range.
For capturing the electro-optic response and bandwidth of PN microheaters, which requires operation at higher speeds, a high-speed photodetector (Newport 1544-A) with a bandwidth of up to 12 GHz and a vector network analyzer (Siglent SVA1075X) with a bandwidth between 100 KHz and 7.5 GHz were utilized. In these measurements, a bias tee (Mini-Circuits ZFBT-4R2GW+) was used to provide a DC offset to the RF signal coming from the VNA.
C. Switching Phase-Change Materials
After sputtering the PCMs onto the chip, they were in an amorphous state. To crystallize the PCMs, an annealing process was used in which the chip was heated to 200°C on a hot plate for 10 min. After initial crystallization, amorphization and recrystallization were achieved by sending electrical pulses. For the experiments in Figs. 2 and 3, we used a 100 µs amorphization pulse with amplitudes ranging from 5.6 V to 6.4 V, and for recrystallization, a 10 ms pulse with 1 ms rise and fall times and amplitudes ranging from 3.2 V to 4 V was used. Crystallization pulses were generated using a Rigol DG4102 arbitrary waveform generator, while a 50 MHz Wavetek pulse generator was used to provide the higher voltage pulses for amorphization. A relatively high voltage was required in these experiments due to the limited current output of the pulse generator.
For the experiments in Fig. 5, we replaced the analog pulse generator with the programmable Rigol DG4102 arbitrary waveform generator and high-speed current amplifier (50Ω output impedance), allowing more accurate control over the pulse parameters. By optimizing the pulse shape, we were able to achieve reversible switching using a 5 µs amorphization pulse with amplitudes ranging from 3.63 V to 4.44 V (amorphization energy of 315 nJ to 472 nJ) and a 5 ms, 2.8 V pulse for full recrystallization (crystallization energy of 168 µJ). Note that these reported voltages are measured across the device and correspond to a 4.5 V to 5.5 V pulse for amorphization and 3.4 V pulse for crystallization when directly measuring the output of the current amplifier on a high impedance load. To switch between nonvolatile programming with the waveform generator and volatile reverse-bias sweep with a Keithley 2400 [see Figs. 5(a) and 5(b)], we used an HP 3488A switch control unit with RF multiplexer card.