Ultrafast optical circuit switching for data centers using integrated soliton microcombs

The study addresses the challenge of scaling data center networks as electrical packet switches face limits due to the slowdown of Moore’s law. It investigates whether nanosecond-scale wavelength switching can enable optical circuit switching (OCS) with high bandwidth, low latency, and improved energy efficiency. The authors propose a disaggregated tunable transceiver architecture that decouples wavelength generation from wavelength selection, using an integrated Si3N4 soliton microcomb as a multi-wavelength source and SOAs as fast, low-loss switching elements. The purpose is to demonstrate sub-nanosecond wavelength switching with practical data transmission formats (25 Gbps NRZ and 50 Gbps PAM-4) and to validate a compact PIC-based wavelength selector (AWG + SOAs) suitable for scalable, power-efficient data center deployments. The study highlights the importance of achieving reconfiguration overheads well below 10% for small-packet workloads, motivating sub-ns switching to keep up with 50 Gbps-class links.

Prior work includes MEMS-based OCS with high port counts (~1000) but millisecond switching times that underutilize bandwidth. Faster photonic switching has been shown using SOAs, EAMs, and MZIs with nanosecond performance. AWGR-based OCS with tunable lasers provides a passive, fault-tolerant core but conventional tunable lasers have tuning latencies of ~10 ns or higher; custom solutions reach ~5 ns yet remain marginal for stringent targets. A disaggregated laser architecture decoupling wavelength generation from selection can achieve sub-ns tuning, previously demonstrated with banks of lasers and SOA/AWG selectors. Integrated Si3N4 soliton microcombs have emerged as broadband, power-efficient, and grid-aligned multi-wavelength sources used in ranging, microwave photonics, OCT, and coherent communications, with recent advances enabling on-chip or electrically pumped operation. Compared to laser banks, microcombs offer dense carriers across C/L bands, precise ITU grid spacing (50/100 GHz) without extra guard bands, and can be shared across many nodes, amortizing power while simplifying stabilization and fast switching control.

Architecture and experimental approach: The authors implement a disaggregated tunable transceiver where a packaged high-Q Si3N4 microresonator generates a single-soliton microcomb (FSR ~99.5 GHz) pumped by a CW laser amplified by an EDFA. Soliton initiation uses forward tuning followed by backward switching; the strong pump is filtered via OADM. The soliton is post-amplified (low-noise EDFA) and demultiplexed by a 48-channel 100-GHz AWG (C-band). Individual comb lines are switched either with discrete SOAs or with an integrated Indium phosphide PIC comprising an AWG and a bank of SOAs functioning as a reflective wavelength selector (one facet high-reflection coated so the single AWG serves as both MUX/DEMUX). Alignment of comb lines to PIC AWG is achieved via chip temperature tuning. Device fabrication and characteristics: Si3N4 microresonators are fabricated using the photonic Damascene reflow process yielding ultra-smooth sidewalls and ~1 dB/m linear loss, with intrinsic Q0 > 15 million, intrinsic linewidth ~15 MHz, FSR 99.5 GHz, and waveguide cross-section 1500 × 900 nm^2. Double inverse nanotapers are used for coupling; packaged fiber-chip-fiber coupling efficiency ~15% due to UHNA-to-SMF-28 splice losses (~2 dB). Power conversion from CW pump to single soliton is ~2%. Switching control: A custom switching control unit supplies bias currents and high-speed electrical signals to SOAs with <100 ps time synchronization accuracy between control and data. To minimize dead time, the next channel begins turning on while the current channel is turning off, which degrades the apparent ER during switching because signals sum; zero-level current is ~1 µA. Data transmission setup: For burst-mode demonstrations, four switched comb carriers are further amplified to overcome ~7 dB MZM insertion loss (20 GHz LiNbO3 MZM at quadrature). A wide-band OBF (~20 nm) suppresses ASE from SOAs/EDFA and rejects next-FSR AWG leakage. Burst sequences at 25 GBd NRZ (2^15 PRBS) and 50 GBd PAM-4 (2^16) are generated by an AWG/DAC, detected by a 50-GHz PD, amplified by a TIA, and captured by a real-time oscilloscope (160 GSa/s). Offline DSP per prior work recovers data and estimates BER while varying received optical power via VOA. PIC-based selector characterization: The InP PIC integrates 23 SOAs (19 for switching, 4 references) connected to a 1×32 AWG (50 GHz spacing). One facet is HR-coated to reflect selected channels back through the AWG and SOAs to the input waveguide, enabling compact reflective operation. Comb-to-AWG alignment is tuned thermally, resulting in seven aligned comb lines during tests. High-speed RF probes contact on-chip electrodes; observed overshoot arises from impedance mismatch and can be mitigated by drive optimization.

- Sub-nanosecond switching with discrete SOAs: For a single comb channel (AWG CH 37, 1554.9 nm), measured 10–90% rise and fall times of 493 ps and 395 ps at 120 mA SOA current. More than 20 individual C-band comb channels (1540–1564 nm) showed <520 ps switching. - Multi-channel switching: Demonstrated sub-ns switching among four comb channels separated by up to ~5.6 nm (also 0.8–4.9 nm spacings shown), using a 2.56 ns guard zone to ensure smooth transitions. AWG provides ~30 dB isolation (discrete setup); OBF limited maximum channel separation. - Burst-mode data transmission with discrete SOAs: While switching among four channels, achieved BER below 5×10^-5 (KP4 FEC threshold) at ROP ≈ −12 dBm for 25 Gbps NRZ and ≈ −8 dBm for 50 Gbps PAM-4. PAM-4 showed an error floor emerging for ROP > −6 dBm due to ASE. - PIC-based AWG+SOA wavelength selector: >20 dB adjacent-channel isolation. Thermal alignment yielded seven matched comb channels. Measured 10–90% switching times between two channels ranged from ~375 ps (min) to ~820 ps (max). Overshoot observed due to RF impedance mismatch. - PIC-based burst-mode NRZ: Achieved BER below FEC threshold when switching between two channels at ROP > −11 dBm. Crosstalk asymmetry affected performance: channel 41 outperformed channel 40 due to ~8 dB lower crosstalk from 41→40 than 40→41. - Scalability: With 40 wavelengths per transmitter and 48 50-Gbps channels (e.g., NVIDIA A100), up to 1920 nodes can be directly addressed; with 256 uplinks in modern ToR switches, up to 25,600 racks could be interconnected via wavelength assignment. - Power analysis and sharing: Current multi-wavelength source consumes ~30 W total electrical power, providing >60 carriers with per-line optical power > −20 dBm (~500 mW electrical per carrier). With improved coupling, on-chip actuators, and efficient DFB pump, projected consumption <193 mW per carrier (~15.5 W total). Dispersion-optimized microresonators can generate up to 122 lines > −14 dBm without post-amplification; separate C/L-band amplification could yield ~13 dBm per line. Sharing a central comb across 32 racks with splitter-amplifier hierarchy provides ~−4 dBm per line at ~2.57 W electrical per rack using a commercial EDFA (or ~1.115 W with on-chip amplifiers).

The results demonstrate that fast, sub-nanosecond wavelength selection is feasible using an integrated soliton microcomb source with SOA-based selectors, meeting the stringent reconfiguration requirements for small-packet, high-speed data center workloads. By decoupling wavelength generation (shared microcomb) from selection (per-rack SOA/AWG), the architecture maintains a passive network core (AWGR) for fault tolerance and scalability while providing per-rack ultrafast switching without complex laser tuning. The BER performance under burst-mode NRZ and PAM-4 during active switching validates practical data transport. The PIC-based reflective AWG+SOA selector confirms that compact, integrated switching elements can retain sub-ns dynamics and sufficient isolation, with further room to reduce impedance-induced overshoot and crosstalk. From a systems perspective, a shared comb amortizes source power across many racks, making overall efficiency approach that of the amplifiers. The architecture supports large-scale interconnects by leveraging many wavelengths, with detour routing enabling full-bandwidth communication without centralized scheduling, at the cost of up to 2× throughput overhead that can be offset with additional uplink bandwidth. These findings collectively indicate a viable path toward flat, energy-efficient data center networks with nanosecond OCS in the post-Moore’s law era.

The paper presents a disaggregated, fast-tunable transceiver architecture using an integrated Si3N4 soliton microcomb as a shared multi-wavelength source and SOA-based wavelength selectors for ultrafast OCS. It experimentally demonstrates: (i) >20 C-band comb channels switched in <520 ps with discrete SOAs; (ii) burst-mode 25 Gbps NRZ and 50 Gbps PAM-4 transmission during four-channel switching with BER below FEC thresholds; and (iii) a compact InP PIC (reflective AWG + SOAs) achieving sub-ns switching (375–820 ps) and 25 Gbps NRZ with BER below FEC at ROP > −11 dBm. The architecture scales to thousands of nodes/racks and benefits from sharing a high-quality comb source to improve power efficiency. Future work should focus on improving fiber-chip coupling (reducing splice loss), integrating low-power on-chip actuators for soliton control, optimizing AWG crosstalk and insertion loss, mitigating RF impedance mismatch in SOA drives, and employing lower-FSR (25/50 GHz) microcombs and higher-output comb configurations to support denser channels and higher-order modulation with improved margins.

- ASE noise limits: PAM-4 exhibited an error floor for ROP > −6 dBm due to ASE from SOAs/EDFAs. - Channel separation constraints: Maximum tested separations were limited by the available optical bandpass filter; L-band channels were not tested due to lack of an L-band AWG. - Coupling and packaging losses: Overall coupling efficiency ~15% with ~2 dB UHNA–SMF-28 splice loss; limits output power and PAM-4 performance. - PIC crosstalk and insertion loss: >20 dB isolation achieved, but crosstalk asymmetry impacted BER; insertion loss could be reduced with improved AWG design. - Drive signal integrity: Overshoot during on-chip SOA switching due to RF impedance mismatch; requires electrode/driver co-design and matching. - Power consumption in current setup: Present source consumes ~30 W; while sharing amortizes power, further integration and component improvements are needed to reach projected per-carrier efficiencies.