20 GHz Silicon Integrated Optical Ternary Content Addressable Memory (CAM) Cell

We propose and experimentally demonstrate an optical ternary content addressable memory cell operating at a record-high search speed of 20 Gb/s on a silicon photonic coherent Crossbar array with an energy efficiency of 0.2 pJ/bit.


Introduction
The rapid advances of networking equipment, along with the rise of Artificial Intelligence and neuromorphic-based computing applications, have stimulated a gradual shift towards more memory-centric and application-specific compute engines in modern Data Centers (DC).These aim to undertake data-intensive operations [1], such as address look-up [2] or packet inspection [3], for accelerating packet-switching functionalities.The look-up and forwarding information base entries of packet switching and routing engines are performed in special content addressable memory (CAM) tables that rely on 2D configurations [2], [4], allowing in this way a fast parallel search within a single clock cycle.Yet, look-up of the optical packet continues to be performed in the electronic domain via electronic CAM banks that can hardly reach 1 GHz processing rates, thus failing to keep up with the frantic scaling of optical line-rates and forming a constantly growing gap between transmission and forwarding speeds.
In order to efficiently resolve packet addressing, optics have recently started to penetrate the look-up address domain, by developing optical CAM alternatives that can be utilized in fast look-up tables [5]- [9].Early experimental demonstrations of InP SOA-MZI-based single optical binary [5] and ternary CAM (TCAM) cells [6] supporting 10 Gb/s search speeds have been presented and more recently extended to a complete 4×2-bit CAM address look-up prototype [7], requiring, however, a rather high power consumption due to the use of semiconductor optical amplifiers (SOAs).An alternative 2-bit CAM Matchline (ML) architecture relying on cascaded integrated silicon photonic (SiPho) micro-ring resonators (μRR) [4] was also recently demonstrated, necessitating again a wavelength-encoded scheme for the search word bits and performing at a speed of 4 Gb/s.
In this paper, we demonstrate a novel, record-high search speed SiPho TCAM cell that provides search operations at 20 Gb/s.Its architecture extends along the principles of the recently proposed electronic Xbar-based TCAM demonstration [1], utilizing the recently introduced coherent photonic Xbar layout [11] that has been fabricated as a 4×4 silicon photonic chip with Silicon Germanium electro-absorption modulators (SiGe EAMs) as volatile memory cells [12], [13].We present experimental results for a fully functional 1-bit TCAM cell at data-rates ranging from 5 Gb/s up to 20 Gb/s, reporting an energy efficiency of ~207 fJ/bit over a cell real estate of 0.095 mm 2 .

Optical Ternary Content Addressable Memory table
Figure 1(a) presents a conceptual layout of a Binary/Ternary CAM table architecture, for the ML operation of a 4-bit search address word.The latter is broadcasted to the stored words of the TCAM table which takes the decision on which port to forward the packet to, by generating a "match" signal only at the TCAM row which has an identical stored content.The blue-highlighted area includes the CAM table entry comparison with the respective search word and depicts the implementation under study.
Figure 1(b) illustrates the basic optical TCAM cell architecture and its operational principle for all three cases of a stored bit (S.B.) of "0", "1" and "X" value.Specifically, the search bit along with its complementary bit-bar value are injected as two different yet coherent optical beams into the single-bit TCAM cell.The TCAM cell comprises two optical branches, each one including an intensity modulator (IM) denoted as S1 and S2 for the upper and lower branch, respectively.The two IMs are used for encoding the stored bit information as a binary vector with its value depending on whether each IM operates in its ON or OFF state.The two optical branches are then combined via a 2:1 photonic combiner or coupler that forces the two constituent optical beams to interfere constructively, using phase shifting (PS) elements to ensure the required phase matching.In this way, the cell output carries the dot-product between the stored and search bit binary vectors, implementing the required ML operation.As illustrated in Fig. 1(b), the [ON, OFF] state of the TCAM cell represents the stored word "0", since a [Bit=0,Bit =1] input will provide a zero-level output corresponding to a "match" operation, while if [Bit=1,Bit =0] comprises the search bit, the TCAM cell will output a logical "1"-level indicating a "mismatch".The full 1-bit TCAM truth table is depicted in Fig. 1(c), clearly denoting that matching is obtained only when a TCAM cell output signal equal to the logical "0".This cell architecture can scale to a multi-bit optical TCAM bank by adopting the coherent photonic Xbar architecture proposed in [11], which allows for complete matrix-vector multiplication operations.Figure 1(d) visualizes the proposed multi-bit TCAM implementation over the coherent photonic Crossbar.Specifically, an N-bit TCAM would require the encoding of the N-bits search word along with their complementary bit-bars via a search word optical encoding unit as inputs of a 2Nrows TCAM Crossbar.These are, then, equally broadcasted into the Crossbar columns, via directional couplers, that host all the necessary stored words.The ML operation is then implemented in parallel for all the stored words by effectively multiplying the search-word bit-bar encoded vector with the stored-vector of each column, providing a logical zero output level for a CAM table "match".

Experimental Setup and Results
The experimental validation of the proposed optical 1-bit TCAM was implemented over the 4×4 SiPho Crossbar chip [12], [13] a microscope photo of which is depicted in Fig. 2(a).The SiPho chip was fabricated in imec's ISSIPP50G platform using PDK-ready components and incorporates 50 um-long, 50 GHz SiGe EAMs as search and stored word encoding units, and TO PSs for safeguarding constructive interference.As depicted in the experimental setup shown in Fig. 2(b), two out of four rows and three out of the four columns were utilized, with rows 1 and 2 used for encoding the search Bit − Bit information and columns 1, 2 and 3 used for storing the "1", "0" and "X" stored bit values, respectively.A single continuous wave (CW) optical beam was produced by a tunable laser source (TLS) at   The TCAM output signals emerging at the output of the respective Crossbar column, were forwarded to an EDFA for amplification and were subsequently captured by a 60 GHz photodiode (PD) and a 75 GHz sampling oscilloscope (OSC).All the EAM and PS modules were electrically biased by a DC control plane.Specifically, the EAMs implementing the stored bit of the TCAM cell units were configured to operate at reverse bias voltages between {0, -3V}, while the EAMs implementing the search bit were reversed biased at approximately -1.5 V to ensure higher optical modulation amplitude (OMA) operation.
During the experimental evaluation of the 1-bit TCAM cell we recorded the Crossbar outputs and filtered the high-bandwidth noise using a low-pass Butterworth filter.Figure 3(a) illustrates the captured time traces of the search Bit and Bit values, as well as the traces of the TCAM output contents for the stored bit cases of logical '0', '1' and 'X', at 10 Gb/s, for a time window of 15 bits.The successful ML operation of the optical TCAM is highlighted with the green and red scatter crosses that indicate the "match" and "mismatch" decision of the ML comparison at each logical stored word output column.The eye diagrams of all 80,000 search bits that were sent to the optical TCAM are plotted in Fig. 3(b) next to their corresponding time traces.The respective eye diagrams of the data collected when the TCAM was operated at 5 and 20 Gb/s are depicted in Fig. 3(c), with Fig. 3(d) illustrating the respective Q-factor values that ranged in all cases between [3.5, 5].Finally, the energy efficiency of our 20 Gb/s binary CAM cell was calculated equal to 0.2 pJ/bit, considering the CAM cell EAMs of  = 0.8 / responsivity and -9 dBm of input optical power, as well as the TO PSs requiring Pπ=4 mW electrical power for 0-π phase shift specifications.

Fig. 2 .
Fig. 2. (a) Microscope photo of the photonic 4×4 Crossbar prototype.Inset: Single TCAM cell comprising two EAMs and two TO PSs (b) Experimental setup for the implementation of a 1-bit TCAM ML operation at 5-20 Gb/s.The search word encoding unit, the 4×4 Crossbar and the 3 TCAM cells that were experimentally operated are highlighted with the red, yellow and purple rectangles, respectively.

Fig. 1 .
Fig. 1.(a) Schematic layout of a Binary/Ternary CAM table, (b) Optical Ternary CAM cell implementation for the CAM content cases of '0', '1' and 'X', (c) Truth table of a single-bit Optical Ternary CAM cell and (d) Conceptual N-bit Ternary photonic Crossbar CAM table architecture, following a search word optical encoding unit.
1563 nm and consequently injected to the photonic Crossbar via a fiber array and a TE grating coupler.Bit and Bit NRZ PRBS7 signals representing the search bits were produced by a 38 GHz arbitrary waveform generator (AWG) and after amplification to 1.8 Vpp were injected to the search bit encoding EAMs.The stored bit encoding EAMs of the first three Crossbar columns were set to the ON/OFF states according to the stored bits they were representing.