Introduction

Energy efficient programmability in large-scale photonic circuits is essential for realizing deployment in applications such as hardware accelerators for machine learning to quantum computing. Typical programmable approaches entail thermo-optic (TO)1, free-carrier dispersion2, electro-optic (EO)3, and piezo-electric based mechanical tuning4. These solutions all require large static-power consumption due to their volatile nature. The TO effect requires  >10 s of mWs, whereas the free-carrier dispersion, EO and mechanical effects require a constant voltage bias. As a result, static power consumption will significantly scale with the number of photonic elements and time of operation. Another challenge is being able control the phase errors associated with the fabrication of phase-sensitive devices such as ring resonators, arrayed waveguide gratings (AWGs), lattice filters, etc. These phase errors are dependent on waveguide width, thickness, and refractive index non-homogeneity which can result in wavelength registration changes of ±60 pm5,6. Furthermore, multi-wavelength laser arrays usually require fine wavelength tuning to correct for wavelength errors cause by fabrication imperfections and environmental/device temperature fluctuations7. The volatile phase tuning solution requires constant standby power consumption from an external voltage source and can significantly affect the power budget of the entire optical communications link. Recently, non-volatile tuning has emerged as a potential solution for energy-free static phase tuning. These efforts include the use of chalcogenide phase-change memory (PCM)8, ferroelectrics (BaTiO39, LiNbO310, PZT), floating-gate memory (FGM)11,12, and memristors13,14,15,16,17,18,19,20. In regards to non-volatile wavelength tuning of semiconductor lasers, our group has demonstrated memristive based solutions20. As an alternative to memristors, and to address the issues of reliability and manufacturability, we have investigated the use of charge-trap memory (CTM) or FGM based solution for non-volatile wavelength tuning12,21,22.

In this paper, we explore a multi-bit optical memory array based on heterogeneous III-V/Si micro-ring resonator lasers (MRL) with co-integrated charge-trap memory (CTM). The array consists of 5 cascaded MRLs, each capable of 4 programmable non-volatile states, thus yielding 1024 unique states altogether. Each MRL exhibits full write/erase operation (100 cycles) with wavelength shifts of Δλnon-volatile ~ 80 pm and a dynamic power consumption  <400 pW. Multi-bit write operation (2 bits) is demonstrated for each MRL and verified over a time duration of 24 h. Table 1 shows a list of current state-of-the-art silicon photonic CTM/FGM devices, performance, and the work presented in this paper. Furthermore, we discuss the use of these non-volatile light sources for optical ternary content-addressable memories (O-TCAM) which find utility in memory search 23,24,25.

Table 1 Complete Survey of State-of-the-Art CTM/FGM Photonic Devices
Full size table

Results

Device design

The fabricated MRL devices with co-integrated charge trap memory defined by the III-V/Al2O3/Si region are shown in Fig. 1. Fabrication first starts with a 100 mm silicon-on-insulator (SOI) wafer which consists of a 300 nm thick top silicon layer as shown in Fig. 1a, b. The silicon waveguides are patterned with a deep-UV ASML stepper and then etched with Cl2-based gas chemistry. Next, 10 nm of Al2O3 is deposited on both sides of the patterned silicon and pristine III-V epitaxial wafer via atomic layer deposition (ALD). Next, the III–V epitaxial stack is then wafer-bonded onto the patterned silicon wafer with the an intermediary Al2O3 layer as shown in Fig. 1c. A more detailed description of the fabrication process can be found in the Supplementary section 5.

Fig. 1: Images of fabricated non-volatile MRLs.
figure 1

a Top-view7. b Cross-sectional SEM7. c TEM image of Al2O3 dielectric charge trap layer.d Microscope image of MRL array.

Full size image

The silicon waveguide is defined by a width, height, and etch depth of 1.5 μm, 300 nm, and 245 nm, respectively as shown in Fig. 1b. The wafer-bonded III-V region consist of a p-i-n epitaxial stack and the 3 μm wide active gain region is defined by multiple-quantum wells (MQW) which consists of 6 compressively strained (+0.8%) In0.86Ga0.14As0.55P0.45 well layers surrounded by 7 tensile strained (−0.19%) In0.84Ga0.16As0.29P0.71 barrier layers. An electron stop layer (In0.5Al0.5As) is used above the MQW region for improved thermal performance. The top mesa consists of a 1.5 μm thick p-InP followed by a 200 nm p-In0.53Ga0.47As contact layer. The full epitaxial stack can be found in Supplementary Section 2. A full silicon etch is partially used underneath the n-InP region such that there is electrical isolation between the n-InP and the p++ silicon contact as shown in Supplementary Fig. 1a. The simulated transverse electric (TE) mode is shown in Supplementary Fig. 1b. The 20 nm thick Al2O3 has an experimentally measured refractive index of \({n}_{{{\mbox{Al}}}_{2}{{\mbox{O}}}_{3}}\) = 1.775 which results in a calculated MQW confinement factor of ΓMQW = 5.5089 % with an overall effective index of neff = 3.2505 and group index of ng = 3.4942. The III-V/Al2O3/Si stack was initially designed for volatile phase tuning5,6,26, where in forward bias, an accumulation of carriers exist at the semiconductor/insulator interface. The change in material index Δn and free carrier absorption ΔαFCA is thus due to a voltage-dependent plasma dispersion effect and can be described by the classical Drude model as shown in Supplementary Fig. 1c.

Origins of non-volatile wavelength tuning: charge-trapping mechanisms

Aside from volatile wavelength tuning, non-volatile tuning is possible due to time dependent carrier traps that exist at the imperfect interface of Al2O3/p-Si and Al2O3/n-InP. As shown in Figs. 1c, 2a, the imperfect interface between crystalline silicon and amorphous Al2O3 as well as crystalline n-InP results in the presence of dangling bond carrier traps. These traps can serve to alter the internal fields of the n-InP/Al2O3/p-Si interface and induce non-volatile plasma dispersion based wavelength shifts. For instance, if electron traps exist at the interface of Al2O3/p-Si, the trapped electrons will attract holes towards the interface on the side of the p-Si as shown in Fig. 2a. Likewise, if hole traps exist at the interface of Al2O3/n-InP, the trapped holes will attract electrons towards the interface on the side of the n-InP region. The combined trapping effects of the n-InP/Al2O3/p-Si interface results in a non-volatile plasma dispersion effect that permanently alters the refractive index of the particular region and is the origin of non-volatile wavelength tuning. The trapping is a transient behavior and requires appropriately long programming and reset voltage biases as described by the Heiman model11. This transient behavior can be described by an occupation probability differential equation as shown in Eq. (1).

$$\frac{d{F}_{tA}^{x}}{dt}={v}_{P}^{x}{\sigma }_{P}^{x}\left[P(1-{F}_{tA}^{x})-{F}_{deg}^{x}{F}_{tA}^{x}{n}_{i}^{x}exp\left(\frac{{E}_{tA}^{x}-{E}_{i}^{x}}{{k}_{B}T}\right)\right]\\ -{v}_{N}^{x}{\sigma }_{N}^{x}\left[N{F}_{tA}^{x}-\frac{\left(1-{F}_{tA}^{x}{n}_{i}^{x}\right)}{{F}_{deg}^{x}}exp\left(\frac{{E}_{i}^{x}-{E}_{tA}^{x}}{{k}_{B}T}\right)\right],$$
(1)
$${\sigma }_{P}^{x}(d)={\sigma }_{P}^{x}ex{p}^{-{\kappa }_{e}d},{\sigma }_{N}^{x}(d)={\sigma }_{N}^{x}ex{p}^{-{\kappa }_{h}d},$$
(2)
$${\kappa }_{h}^{2}=\frac{2{m}_{h}^{*}\left({E}_{tA}^{x}-{E}_{v}^{x}\right)}{{\hslash }^{2}},{\kappa }_{e}^{2}=\frac{2{m}_{e}^{*}\left({E}_{c}^{x}-{E}_{tA}^{x}\right)}{{\hslash }^{2}}.$$
(3)

where \({F}_{tD}^{x}\), \({v}_{P}^{x}\), and \({v}_{N}^{x}\) represent the carrier trap occupation probability, hole thermal velocity, and electron thermal velocity respectively. x represents either the Al2O3/p-Si or Al2O3/n-InP interface. \({\sigma }_{P}^{x}\) and \({\sigma }_{N}^{x}\), defined by Eq. (2), denote the carrier trap cross-section for holes and electrons, P and N are the free hole and electron concentrations, \({F}_{deg}^{x}\) is the degeneracy factor, \({n}_{i}^{x}\) is the intrinsic carrier concentration, \({E}_{tD}^{x}\) and \({E}_{i}^{x}\) are the energy levels of the carrier traps and intrinsic energy levels. The four terms on the right hand side of Eq. (1) consists of four distinct rates: 1) a trapping rate by conduction band electrons, 2) a discharge rate to the conduction band, 3) a discharge rate due to valence band hole capture, and 4) a trapping rate due to emission of holes to the valence band. These rates depend linearly on the carrier trap cross-sections, and which rate dominates is determined by the carrier surface densities. The carrier tunneling to trap states \({E}_{tA}^{x}\) from the conduction band edge \({E}_{c}^{x}\) is modeled by the term with the electron evanescent wavevector κe as shown in Eq. (2)–(3). Likewise, tunneling from the valance band edge \({E}_{v}^{x}\) to trap state \({E}_{tA}^{x}\) occurs through the wavevector κh. The mono-energetic trap density is assumed to be uniform up to a specified depth d, and zero beyond that11. 

Fig. 2: Charge-trapping mechanisms.
figure 2

a SEM image of n-InP/Al2O3/p-Si interfaces and regions of carrier traps. b Energy band diagram of initial state and during program operation. Calculated carrier-trap occupation probability for (c) Al2O3/p-Si, and (d) Al2O3/n-InP interfaces.

Full size image

Figure 2c–d shows the calculated carrier-trap occupation probabilities for both interfaces of the n-InP/Al2O3/p-Si stack. Without any bias, the “initial” state has a value of zero occupation probability. After programming with a voltage bias of sufficient time duration, the occupation probability reaches a saturated value which becomes the on-set of non-volatility. By tuning off the bias, there exists a permanent occupation probability over time, indicating non-volatility. Erasure of this non-volatile state is achieved by applying a bias of opposite polarity to sweep out trapped carriers back towards the semiconductor regions. The material parameters used in the temporal occupation probability calculations are itemized in the Supplementary Section 2.

Device characterization

Initial light output vs. current–voltage measurements (LIV) were performed on the left and right output facets of the MRL laser as shown in Supplementary Fig. 2a with an observed threshold current of approximately 12 mA. The experimental details are discussed in Supplementary Section 3. This particular ring has a silicon waveguide radius of 23.62 μm and results in a calculated free spectral range (FSR) of 3.515 nm defined by FSR = λ2/ngL. The actual measured FSR is closer to 3.167 nm (Supplementary Fig. 2b), possibly due to variability of the hybrid III-V/Si optical mode and fabrication errors. Significant multi-mode and bi-directional lasing exists due to the presence of “kinks” in the LIV plots. It has been shown that back-reflections from imperfect sidewall etching can cause the enhancement of intra-cavity back-reflections which degrade stable lasing and can trigger frequent laser direction changes27. This is due to the strong coupling between CW and CCW modes which manifests itself as fluctuations in the LIV curve. This issue can be remedied by forcing the laser to operate in a uni-directional manner. One such strategy is to introduce self-injection locking with the use of an S- shaped intra-cavity28. Along with the LIV plots, we simultaneously tracked the optical spectrum of both left and right outputs shown in Supplementary Fig. 2b-c. These plots are useful for mapping out regions of single-mode operation and either clock-wise (CW) or counter-clockwise (CCW) operation. Single-mode operation is possible at certain injection current values due to one cavity mode dominating the stimulated emission by being responsible for depleting the majority of electronic carriers29. For our case, we decided to use an injection current of 19 mA to achieve simultaneous single longitudinal-mode operation with nearly equal CW and CCW power.

Supplementary Fig. 2d–e illustrates the volatile and non-volatile wavelength shifts that exist during the programming process. Supplementary Fig. 2e shows the normalized shifts vs. programming time. By applying a “write” operation of 0  → −6 V for 20 min, a weak red-shift occurs due to an enlarging of the depletion region on both sides of the Al2O3/Si and Al2O3/III-V interfaces. Next, by turning off the applied bias (−6  → 0 V), a non-volatile blue shift of  ~50 pm is observed. In order to “erase” this non-volatile state, a bias of 0  → +6 V is applied for 20 min. During this erase process, a large volatile blue-shift of  ~100 pm is observed in accordance with carrier accumulation at the interfaces of Al2O3/Si and Al2O3/III-V. Finally, a full ‘reset’ back to the original laser wavelength is achieved by turning off the bias. This entire procedure is illustrated in Supplementary Fig. 2e with the tracked optical spectra shown in Supplementary Fig. 2d. The power consumption is tracked throughout the entire process and is approximately 100s of pWs during “write” and ‘erase’ operation as shown in Supplementary Fig. 2f. The non-volatile state consumes 0 static power.

In order to determine how robust the non-volatile wavelength shift is, we performed cyclability tests as shown in Fig. 3. The same programming operation (Supplementary Section 3) was used with the exception that the “write” and “erase” times were shortened to 5 min. The non-volatile and reset spectra for the left output were recorded for 100 cycles as shown in Fig. 3a, b. Likewise, the data for the right output was recorded simultaneously and shown in Fig. 3d, e. By tracking the peak wavelengths of the non-volatile/ reset states for the left and right outputs, it is observed that non-volatile shifts of  ~80 pm (Δneff,non−volatile ~  2 × 10−4) is possible as shown in Fig. 3c. With the increased amount of cycles, there is an overall blue-shift of the entire spectrum, thus indicating the possibility of material degradation. The extinction ratio (ER) between non-volatile and reset states for both outputs are shown in Fig. 3f. We suspect, the 40 dB ER for the left output is an lower limit because the laser line-width measurement is limited by the optical spectrum analyzer (OSA). The right output has a decreased ER mainly due to the weaker power coming from the CCW direction. We performed measurements across 4 dies spaced 10 mm apart to determine device uniformity. LIV measurements of one MRL from each die indicate threshold currents of 7.7, 10.0, 7.3, and 16.5 mA. The variability in the threshold current indicates either non-uniform gain/loss within the MRL structure possibly due to non-uniform etching or lithography of the III-V mesa and silicon waveguide. We have also performed non-volatile write operation (0 V  → −6 V) and recorded non-volatile wavelength shifts (−6 V  → 0 V). The mean non-volatile wavelength shift from die 1–4 was measured to be −79.0, −82.6, −86.8, and −89.8 pm respectively. The standard deviation of was determined to be 3.54, 1.91, 4.02, and 2.22 pm respectively. This indicates that although the gain/loss have variability, the non-volatile shifts due to charge trapping in the oxide is somewhat repeatable according to the statistics above.

Fig. 3: Non-volatile MRL cyclability tests.
figure 3

a Optical spectrum of 100 non-volatile cycles for left output. b Corresponding 2-D optical spectrum map. c Peak wavelength tracking for left and right outputs during non-volatile programming. d Optical spectrum of 100 non-volatile cycles for right output. e Corresponding 2-D optical spectrum map. f Extinction ratio (ER) of left and right output.

Full size image

Figure 4a shows a schematic of 5 MRLs cascaded onto a common bus waveguide with left and right output ports. MRL1 has a radius of 24.25 μm and each successively cascaded MRL consists of an incremental length of 243 nm to yield a calculated wavelength channel separation of 2 nm. The LIV and corresponding optical spectrum of each MRL is shown in Fig. 4b, c. As expected, without any intra-cavity filters, each of the MRLs exhibit multi-modal and bi-directional lasing as seen in LIV kinks. The optical spectra for the left and right outputs are plotted such that a proper bias current of 27 mA is chosen to yield single-mode lasing in the CW and CCW direction.

Fig. 4: 5 channel MRL array characterization.
figure 4

a Schematic of the MRL array. b MRL LIV measurements for left and right outputs. c Corresponding optical spectrum. d Measured optical spectra of non-volatile lasing states for all 5 MRLs with time duration data overlapped. e Close-up of non-volatile lasing states for each MRL. f Tracked peak wavelength of the various non-volatile states.

Full size image

Next, we apply non-volatile programming voltages of 0, −2, −4, and −6 V to each MRL for 5 minutes respectively. Time retention of the non-volatile state is determined by recording data after the voltage source is turned off. This results in 4 non-volatile laser states for each MRL. Fig. 4d shows the non-volatile spectrum of all 5 MRLs with overlapped time duration. Fig. 4e illustrates a zoomed-in snapshot of each non-volatile lasing state. Taking into account 5 MRLs, this will yield an optical memory bank of 1024 distinct states. By tracking the peak wavelength of each non-volatile laser state, it is determined that a non-volatile wavelength shift of −80 pm is possible. It is interesting to note that beyond −6 V, the Al2O3 enters a memristive state where filamentation occurs causing even larger wavelength shifts 13.

Future improvements in the magnitude of non-volatile wavelength shifts can be achieved with using multi-layer CTM or floating gate memory structures11,12,30 and is detailed in Supplementary Section 6. These types of structures consist of alternating layers of high-k dielectric materials such as n-InP/Al2O3/HfO2/SiO2/Si. The Al2O3 and SiO2 serve as the tunneling/blocking oxide while the HfO2 acts as the charge trap region. In this case, the calculated non-volatile wavelength shift is  ~1.6 nm (Δneff,non−volatile ~ 40 × 10−4), a significant improvement over the structure in this work. Although, the non-volatile switching speeds in our work is slow, there are avenues to improve this. For instance, It is known that CTM switching speeds can be improved with various rapid thermal annealing (RTA) conditions based on temperature and time. These programming speeds can be improved from milli-second to mico-second ranges based on RTA annealing at 450 °C for a time duration of 15 min31,32. Future work will explore the use of the RTA technique prior to wafer-bonding to achieve improvements in programming speeds33. Future work on increasing charge trap sites will be explored via increasing oxygen vacancies and defect impurities using a variety of methods 34,35,36,37.

Non-volatile micro-ring laser content-addressable memory

We found that the programmable memory feature of the MRL, introduced in this work, is useful in implementing a compact optical ternary content-addressable memory (O-TCAM). Content-addressable memory (CAM) is a special type of memory designed to search its entire content within a single clock cycle, indicate if any of the stored Data words match the word of interest, and return the match address. This high-speed table search functionality well-suits routers, switches, and SmartNICs, which are required to quickly execute packet routing38 and inspection39. Recently it has been demonstrated that ternary CAMs (TCAMs) with encoded Search and Data words can execute in-memory computing functions, necessary for various machine learning algorithms40. The Search and Data words in TCAMs can be in one of three states—0, 1, or X (don’t care), resulting in improved search flexibility compared to binary CAMs (BCAMs). Optical CAMs has an advantage over electrical CAMs in operating at signal data rate23,41, thus speeding up the search with reduced latency. In our previous work, we developed two types of micro-ring-based O-TCAM architectures—wavelength-division multiplexing (WDM) O-TCAM and time-division multiplexing (TDM) O-TCAM23. Both architectures employ a special dot-product encoding scheme40, which we found to be suitable for the implementation of microring-based O-TCAMs that capable of measuring Hamming distance, an important metric for similarity learning. To support the three CAM states, each Search and Data symbol consists of two bits. In the case of the WDM O-TCAM this encoding scheme requires two micro-ring modulators (MRMs) for the electro-optic conversion of the ternary Search symbol and two micro-ring resonators (MRRs) to store the ternary Data symbol and compare it with the corresponding Search symbol. This requires twice the number of micro-rings compared to binary search. It is worth noting that in the case of micro-ring-based TDM O-TCAM, binary search and ternary search require the same micro-ring and symbol duration resources.

We propose here a WDM O-TCAM architecture, which utilizes the MRL with integrated non-volatile memory, that reduces the micro-ring count by a factor of two compared to our previous WDM O-TCAM design. The schematic of our MRL-based WDM O-TCAM is shown in Fig. 5. The l-th MRL and MRM, l ∈ {1, K}, have their own set of wavelengths, each indicates on a different symbol state. The MRL state, which is linked to the stored Data symbol, depends on the laser wavelength and the MRM resonance controlled at data rate by the Search symbol. The memory bank is designed for long-term storage, therefore low-speed drivers are sufficient to store the Data word in the MRLs. The encoding scheme used in our MRL-based WDM O-TCAM is described in Table 2 for the l-th ternary symbol. Our encoding scheme is a modification and extension of the one proposed in ref. 42. to suit our MRLs with programmable memory functionality and support ternary search and storage operation. A Mismatch occurs when the light from the MRL is coupled to the MRM, resulting in a current measured at the output of the PD, which is embedded in the MRM. Our MRL-based WDM O-TCAM can be adapted to monitor the Match state at the through port, reducing the PD count to one, similarly to the WDM O-TCAM architecture developed in23. To do so, two modifications of the encoding scheme are required—1) interchanging between the MRM resonances of the Search symbols 0 and 1, and 2) dynamically alternating the MRM between high-Q, to filter either λl,0 or λl,1, and low-Q, to filter both wavelengths. The low-Q mode is designed to block the light at the through port to represents a Search symbol of X (don’t care). This Q switch between the different ternary Search states is possible by modifying the MOSCAP BIAS, however, it is a lengthy update process that results in a slower search speed.

Fig. 5: A schematic of the MRL-based WDM O-TCAM.
figure 5

Each of the Data symbols is stored in a MRL and being compared with the Search symbols at the MRR-based engine. The PD output current indicates on the Match state between the stored Data and Search symbols. Photos of our MRL and MRM chips are shown in the insets.

Full size image
Table 2 The encoding scheme for the l-th symbol in the proposed MRL-based WDM O-TCAM (internal PD / drop port monitoring)
Full size table

To confirm our O-TCAM architecture works, we experimentally demonstrated a 1-symbol MRL-based WDM O-TCAM circuit. We used a micro-ring laser based on quantum dots that had a lasing wavelength of 1310.09 nm and was injection-locked with an external tunable laser to amplify the lasing signal as shown in Supplementary Section 4.2. The lasing signal was then further amplified such that optical power of  ~0 dBm was fed into the MRR array. The MRR is a carrier injection device with a diameter of 19.9 μm and a measured Q of  ~15 K. By applying a modulated voltage bias (i.e., search signal, s(t)) generated from the arbitrary waveform generator (AWG), we are able to observe the match signals of 0 and 1 via the through port of the MRR. Search signal modulation speeds were performed at 50 Mbps and limited by mechanical constraints of a high-speed probe and fiber array. The encoding scheme (modified for through port monitoring) and the engine and memory settings, used in the experiment, are described in Supplementary Section 4.3. Since the microring has fixed Q value and the match signal is monitored at its through port, the search signal is taken from binary set. The search signal, s(t), is shown Fig. 6a and the match signals are shown in Fig. 6b (d = 0), Fig. 6c (d = 1), and Fig. 6d (d = X). The time traces shown in Fig. 6b–d validate that our O-TCAM operates as intended.

Fig. 6: The experimental results of MRL-based WDM O-TCAM architecture.
figure 6

a The search signal, s(t), taken from binary set, which is measured by the real-time oscilloscope. Above the signal, the Search sequence is indicated. b–d The match signal, measured by the real-time oscilloscope, when compared to stored symbol of 0 (subplot b), 1 (subplot c), X (subplot d). The match decision points are marked with a plus symbol, where the Match and Mismatch are indicated by the black and red color, respectively.

Full size image

Furthermore, we simulated in Lumerical INTERCONNECT a 1-symbol MRL-based WDM O-TCAM, which is monitored at the drop port. The MRL is modeled with two continuous wavelength (CW) lasers and the content-comparator engine with a SiPh MRM, which has at its drop port a SiGe avalanche photo-diode (APD) followed by a trans-impedance amplifier (TIA). Both the MRM and APD models are taken from our SiPh process design kit (PDK) that we developed for our research on exascale computing interconnects43,44,45. The light from the CW lasers, embodies the stored Data symbol, is broadcasted to the MRM, which is being modulated by the Search sequence. This mode of operation is equivalent to a comparison between the Data and Search symbols. The simulation settings are summarized in Supplementary Section 6.2. Since the MRM has non-identical insertion loss between the 0 and 1 Search states, the output power of the CW lasers is set to compensate for this difference. To map all content-search combinations, we set the Search sequence to X1X0X and the stored Data to 0, 1, or X. The search signal, driving the MRM at 10 Gbps, is shown Fig. 7a and the match signals are shown in Fig. 7b (d = 0), Fig. 7c (d = 1), and Fig. 7d (d = X). In all three cases, at the decision points (indicated by a plus marker) the match signal follows the decision scheme in Table 2. The eye diagrams of the match signal, when the Search symbol is taken from a binary set, are shown in Fig. 7e (d = 0), Fig. 7f (d = 1), and Fig. 7g (d = X). The open eyes validate that our O-TCAM operates as intended. The high Q value of the MRM (Q = 15 K), which is needed due to the small wavelength shift of our current MRL (80 pm), results in significant spread of the Match/Mismatch pulse beyond its time interval. Similar effect on the match signal was previously observed in23. The interference between the match pulses as well as the contrast between the Match and Mismatch states can be improved significantly with a new MRL design (following the steps described in Section “Device characterization”) that allows higher wavelength shift, leading to better ER and allowing to use MRM with broader bandwidth.

Fig. 7: The simulation results of a MRL-based WDM O-TCAM circuit with a single storing memory and a single engine unit.
figure 7

a The search signal, s(t). The Search sequence is X1X0X. The match signal at the TIA output, when compared to a stored Data symbol of 0 (subplot b), 1 (subplot c), or X (subplot d). The match decision points are marked with a plus symbol, where the Match and Mismatch are indicated by the black and red color, respectively. The eye diagrams at the TIA output, when compared to Data symbol of 0 (subplot e), 1 (subplot f), or X (subplot g). The Search symbol is taken out of a binary set.

Full size image

We further analyzed the energy performance of our MRL-based O-TCAM architecture. The main power consumers are the MRLs, engine-side drivers45, TIAs43, and comparator43,46. The key performance and design rules of our platform are listed in Supplementary Section 7. It is noteworthy to mention that the memory bank drivers negligibly contribute to the total power consumption, because the Data symbols are long-term stored in the MRLs. The impact of each of our O-TCAM devices on the total energy consumption is shown in Fig. 8 for symbol count of 5, 10, and 20. The symbol count is capped at 20 symbols to ensure sufficient channel spacing within the microrings’ free spectral range (FSR), resulting in minimal crosstalk. The required optical power of each of the MRLs was determined through link budget analysis of the O-TCAM circuit. In our platform, an improvement to the link budget is possible by integrating an APD within the MRMs, thus eliminating the drop port losses. The APD sensitivity level corresponds to an NRZ signal with BER of 10−12 43 —these assumptions give us an order of magnitude to the MRL optical power. As our platform capable of heterogeneous integration, we included all necessary photonics devices – light sources, MRMs, and photodetectors—in a single chip. Therefore, the link losses are significantly smaller compared to dis-aggregated circuit designs23, resulting in an average MRL optical power of 14 μW (for 5 symbol O-TCAM circuit). However, due to the low wall-plug efficiency (WPE) of our current devices (0.3%), the impact of the MRLs on the total energy consumption is significant. However, we expect improved WPE up to 20% based on improved waveguide loss, fabrication processes, and past work on efficient quantum dot based MRLs29. This reduces the MRL electrical power by a factor of 66.

Fig. 8: The contributions to energy consumption from the individual components of the O-TCAM with 5, 10, and 20 symbols.
figure 8

The total energy consumption of the MRL-based O-TCAM is indicated by the circle marker.

Full size image

Discussion

In summary, we have demonstrated the co-integration of non-volatile memory elements within III-V/Si light sources. This allows for energy efficient, non-volatile wavelength tuning with 0 static power consumption. An array of 5 cascaded MRLs, representing an optical memory bank, was demonstrated with each laser exhibiting 4 programmable non-volatile states with maximum wavelength shifts of  ~80 pm with  ~40 dB signal extinction ratio. The write/erase operations were repeated for up to 100 cycles with long lived non-volatile time durations lasting up to 24 h. Alternatively, non-volatile wavelength tuning in MRLs find beneficial uses in laser wavelength locking in temperature varying environments. Our past work has shown that lasing wavelengths can red-shift in the range of 0.097–0.17 nm/mA due to current injection heating7. With the improved CTM design in Supplementary Section 6, it is possible to “athermalize" the MRL for current injection values of 9–16 mA with 0 power consumption. We also believe these unique classes of non-volatile III-V/Si lasers, in combination with a seamless heterogeneous photonics platform46,47,48,49,50, can enable applications such as: neuromorphic/brain inspired optical networks, optical phase arrays, tele/data-communications, and future optical near/in-memory accelerator architectures, etc51,52,53,54,55. We explore one use case involving O-TCAM, which is a memory designed to search its entire contents in a single clock cycle. Specifically, we propose a O-TCAM architecture that utilizes cascaded MRLs for a Data word storage as well as cascaded MRMs for match state determination between each of the Search and Data symbols. Because our scheme encodes the ternary Search and Data symbols in the wavelength-___domain, there is a microring resource saving of factor of 2 compared to our previous WDM O-TCAM architecture23. We verified our architecture through simulation and showed that the match signal follows our encoding scheme. We analyzed the energy consumption of our O-TCAM architecture based on our platform key performance and design rules. The heterogeneous integration of both the MRLs and MRMs on the same device minimizes the link losses, such that the required average MRL optical power is 14 μW. The end-to-end energy consumption of MRL-based O-TCAM with 5 ternary symbol is 1156 fJ/sym. By improving the MRLs WPE to 20%, the total energy consumption can be reduced to 703 fJ/sym. While electrical TCAMs integrated in optical networks has lower energy consumption, they require down-converter to accommodate the signal data rate to their search speed. This step is not needed in our MRL-based O-TCAM architecture, therefore resolves the latency issue of content search in current optical network systems.

Methods

Fabrication

In-house device fabrications starts with a 100 mm SOI wafer that consists of a 350 nm thick top silicon layer and a 2 μm buried oxide (BOX) layer. The top silicon was thinned down to 300 nm by thermal oxidation and buffered hydrofluoric (HF) acid etching, thus leaving a clean silicon surface. Silicon waveguides were defined by a deep-UV (248 nm) lithography stepper and boron was implanted to create p++ silicon contacts as shown in Supplementary section 5. Grating couplers, silicon rib waveguides, and vertical out-gassing channels (VOCs) were respectively patterned using the same deep-UV stepper and then subsequently etched 170 nm with Cl2-based gas chemistry. Next, the silicon wafer moved through a Piranha clean followed by buffered hydrofluoric (HF) acid etching to remove any hard masks. Next, an oxygen plasma clean was performed followed by a SC1 and SC2 clean. The III-V wafer is cleansed with acetone, methanol, and IPA, followed by oxygen plasma cleaning and a NH4OH:H2O (1:10) dip for 1 min. Next a dielectric of Al2O3 was deposited onto both III–V and Si wafer via atomic layer deposition (ALD) with a target thickness of 5 nm on each side. ALD deposition temperature happens at 300 °C. The two samples were then mated manually at room temperature and then wafer-bonded under pressure for 300 °C (2 h ramp) for a total of 15 h. After wafer-bonding, the backside of the III–V was mechanically lapped until  ~100 μm of III–V was left. Next, a wet etch was used to remove the remaining InP substrate which stops on the p-contact InGaAs layer. III–V mesas are defined by etching in an Oxford inductively-coupled plasma (ICP) etcher using a Cl2-based gas chemistry stopping right above the active QW regions. The active regions are further defined by an additional lithography step and subsequently wet etched leaving a clean n-contact region consisting of n-InP. A combination of Ge/Au/Ni/Au/Pd/Ti (400/400/240/4000/200/200 Å) was deposited onto the n-InP as an n-contact layer. Metal contact with the p-Si consisted of Ni/Ge/Au/Ni/Au/Ti (50/300/300/200/5000/200 Å). Next, a plasma enhanced chemical vapor deposition (PECVD) SiO2 was deposited for cladding material. This followed up with etched vias. Next, a thick photo-definable BCB layer was used to minimize electrical parasitics. Finally, Ti/Au metal probe pads were defined for final electrical contacts.

Measurement setup for non-volatile MRL testing

The 100 mm wafer is vacuum mounted onto a temperature controlled stainless steel chuck via a semi-automatic probe station. The experimental setup is shown in the Supplementary Section 4.1. All endurance and cyclability measurements of the non-volatile MRL lasers are all automated with a PC. Current injection and MOSCAP phase tuning (volatile and non-volatile) is performed with a Keithley Source Meter (2400). Light is vertically collected from the devices via grating couplers with a 7° polished fiber array. Optical power is measured with Newport power meters (1936-R) and detectors (818-IG-L-FC/DB). Optical spectra was measured using a Yokogawa AQ6370D. Electrical contact on the micro-ring lasers was performed with a 3 contact DC probe where each contact represent current injection, ground, and voltage for phase tuning.

Measurement setup for O-TCAM demonstration

The experimental setup is shown in the Supplementary section 4.2. A 10 % tap is used to observe the optical spectrum and the remaining 90 % is fed into a Thorlabs praseodymium-doped fiber amplifier (PDFA). The amplified signal is then coupled into a micro-ring resonator (MRR) on a separate temperature controlled stage. The stage temperature of the MRL and MRR were set to T = 25 °C. An arbitrary waveform generator (Keysight M8195A) is used to generate the search signal which is then applied to the MRR via a voltage amplifier (SHF M827A) and DC bias offset (Keithley 2400) through a bias-tee. DC probes had to be used for the MRR because high-speed RF probes could not be mechanically aligned due to the close proximity of the optical fiber array. Finally, the match signals are recorded on an Agilent Infiniuum oscilloscope (DSA-X 93204A).