Impact of High-Speed Modulation on the Scalability of Silicon Photonic Interconnects OPTICS 201, March 18 th, Dresden, Germany Meisam Bahadori, Sébastien Rumley,and Keren Bergman Lightwave Research Lab, Columbia University New York, NY, USA Rev PA1 1
Outline Introduction context Silicon Photonics (SiP) Review of ring resonator Low level metrics High level metrics and models for obtaining them Design space exploration and pruning For filtering (modulation), for dropping Power penalties evaluation Identification of ideal ring designs Comparison with existing rings Proposed methodology Conclusion Rev PA1 2
Context Ever larger bandwidths required at all scales From CPU-to-CPU to continent-to-continent Transition to larger bandwidths may occasion shifts to alternative technologies Long distance-link have shifted year ago from copper to optics Such a transition to optics has yet to occur for short distances ( 1-cm) Mainly for practical (=cost) issues: exotic materials required by optical components, unconventional and potentially bulky packages, etc. Silicon photonics can potentially offer a solution to most (many?) of these practical issues Mass and cheap production through CMOS compatibility Close integration with digital logic This does not necessarily mean that shift to silicon photonics will occur Will SiP outperform competing technologies, at bandwidth scale? Is there (semi-hidden) threats to SiP functionality, at architecture scale? Rev PA1 3
Outperforming competitors cost and power Cost: Power: Silicon photonics needs An external laser (array) Edge coupling with external world (at least for laser) Area for driver circuitries (one per wavelength!) To be compared with new cabling solutions, novel signaling schemes used in electronic transceivers Elec. transceivers for intra-chip communication (1cm) achieve 0.1 pj/bit [1] Elec. transceivers for inter-chip communication achieve 5 fj/bit [2] These figures will probably improve by the time silicon photonics reaches maturity We need to target such figures at least Need for a in-depth optimization of device parameters and process Especially at high data-rates [1] Wary et al. High-speed energy-efficient bi-directional transceiver for on-chip global interconnects, 2015 Rev PA1 4 [2] Niitsu, et al. IEEE Trans. On VLSI 20 (7), 2012
Ensuring functionality at scale Very few end-to-end demonstration of silicon photonic systems so far Demonstrations often include tricks Optical amplifiers Piecewise demonstration Loss normalization Device control with sophisticated lab equipment Multiple threats to correct functioning still remain Fabrication variability, susceptibility of components to this variability Insertion losses must be pushed to the minimal and no surprise 5dB can be tolerated Integrated control (i.e. on chip) of devices, area and power it consumes Underestimation of optical impairments as crosstalk Especially at high-rate and at scale [C. Sun et al., Nature 528, 2015] Need for comprehensive models taking into account all these aspects Rev PA1 5
Review of ring resonator n eff Ring: A circular waveguide with properties: R g Effective refractive index n eff Waveguide loss α waveguide (1/m or 1/cm or db/cm) Obtained by numerical methods (FDTD or FEM) Radius R affecting: Resonance λ m = 2π Rn eff m Mode number COMSOL Loss in the ring Gap size affecting: Approx. Function of radius (1/m or 1/cm) Coupling coefficient Rev PA1
Investigating ring loss and coupling coefficients [H. Jayatilleka et al., JLT, 201] Coupled waveguides α ring (db/cm) 40 30 20 4.84e7 R -7.8 + 2 1.04e9 R -.13 + 1.2 Current PDK 0 4 8 12 14 Futuristic PDK Ring radius (µm) κ 0 200 300 400 Rev PA1 7 1 0.5 0 Coupling coefficient R=µm R=µm
High level metrics for ring resonators TR max Ring bandwidth (a.k.a FWHM) Range of filtered frequenties Related to Q-factor Free spectral range (FSR) Transmission at resonant frequency vestiges from filtering BW 3dB ER Transmission at detuning of FSR/2 exactly between resonances TR res FSR Round-trip phase inside the ring Rev PA1 8
Ring resonator filter desired properties Good signal suppression (at resonance) Critical coupling: TR res ~ 0 FSR large enough to maximize WDM capabilities Low suppression outside resonance TR max > -0.05 db (cascading 20 1dB, cascading 0 5dB) Bandwidth large enough to support signal FWHM Ghz > ~Bitrate Gbit OOK Rev PA1 9
Ring resonator filter reducing design space Ring radius (µm) 8 4 Attenuation at resonance - - - -20-20 - -30-30 -40 0 200 300 400 Ring radius (µm) 8 4-0.05-0.1-0.02-0.01-0.001 Minimal attenuation 0 200 300 400 Ring radius (µm) 8 4 0 Ring bandwidth 25 5 0 200 300 400 Ring radius (µm) 0 200 300 400 Rev PA1 8 4
Impact of high-speed signals Ring radius (µm) 8 4 Ring bandwidth @ Gbps 0 25 5 0 200 300 400 Ring radius (µm) 8 4 0 200 300 400 Ring radius (µm) 8 4 Ring bandwidth @ 25Gbps 0 25 5 0 200 300 400 Ring radius (µm) 0 200 300 400 Rev PA1 11 8 4
Ring radius (µm) Impact of WDM and fab PDK 7.5 5.5 5 4.5 Current PDK Futuristic PDK 0 150 200 250 Ring radius (µm) 7.5 5.5 5 Current PDK Futuristic PDK Single channel 4.5 20 channels @ Gbps 4.5 50 channels @ Gbps 0 150 200 250 Ring radius (µm) 7.5 5.5 5 Current PDK Futuristic PDK 0 150 200 250 Zones of feasibility for simple filters Multiple channels: impose additional restriction on insertion loss Insertion loss must stay low (here <0.05dB) around neighboring channel IL(detuning = 50nm/#channels bitrate) < 0.05dB Rev PA1 12
Ring diameter limitation Selection of designs leading to signal suppression of -15dB at least 0 Current PDK Futuristic PDK Below 4µm, loss inflicted to other channels increases sharply Progress on PDK do not provide much help. Min insertion loss (db) -0.5-1 Result in small FSRs of 22 nm at most To be compared with the ~ 200nm -1.5 exploited in the fiber world 3.5 4 4.5 5 5.5 Ring radius Rev PA1 13
Ring resonator drop Third design parameter: output gap size g out R n eff Critical coupling condition: g in Transmission (thru) at resonance: DR max TR max Thru transmission Transmission (drop) at resonance Transmission (thru) out of resonance (at FSR/2): TR res Drop transmission Rev PA1 14
Ring drop parameter space Ring radius (µm) 8 4 0 25 2.5 Ring bandwidth 0 200 300 Ring radius (µm) 8 4-0.03-0.1-0.2-0.5-1 -3 Attenuation (drop) at resonance 0 200 300 Ring radius (µm) 8 4-1 -0.3-0.1-0.03-0.01-0.001 Min attenuation (thru) out of resonance 0 200 300 Ring radius (µm) 0 200 300 Rev PA1 15 8 4-20 Attenuation (thru) at resonance -30-40 -50
Design space pruning Truncation limited Ideal drop loss (0.1 db) But high thru loss (0.05 db) Ring radius (µm) 8 4 Thru loss limited -1-0.03-0.3-0.1-0.1-0.2-0.03-0.01-0.5-1 -0.001-3 Attenuation Min attenuation (drop)(thru) at out resonance of resonance 0 200 300 400 Ideal thru loss (0.002 db) but high drop loss (0.5 db) How to choose? Depend on the architecture (number of thru, drop) number of channels channel rate Drop loss limited Rev PA1 1
Balancing truncation and crosstalk Modulated signals occupy a broader range of bandwidth Relevant part proportional to the bitrate Signal truncation: some of the relevant part is not dropped Occupied spectrum potentially infinite Light intensity T b 2T b 3T b 4T b 5T b time CW laser frequency Light intensity Filtering cross-talk: some of the infinite part of the other channels is dropped as well Truncation comes at the expense of cross-talk and vice-versa Ideal modulator with infinite bandwidth Fall time T b 2T b 3T b 4T b 5T b time 2r b Power Spectral Density Rise time frequency Low bandwidth ring Signal spectrum High bandwidth ring High truncation Very small leakage in other channels Crosstalk if neighboring channel is close Low truncation Rev PA1 17
Bandwidth/Q factor picking 12 Bandwidth 12 Q factor 1500 11 11 Ring radius 9 8 80 Ghz 50 Ghz Ring radius 9 8 Q factor 7 7 0 150 200 250 input gap size Ghz 0 150 200 250 input gap size 2350 Rev PA1 18
Impact of BW/Q: Truncation of the Spectrum Depends on the modulation speed (rate) Depends on the Q of the ring Truncation penalty reflects how much the strength of the information is reduced by the narrow-bend optical filter (db) Power Penalty due to the high Q If Q 0 (infinite bandwidth) TPP 0 (no filtering effect) Taking into account Bandwidth of ring and DATA RATE Resonance freq. of ring (f 0 = c/λ 0 ) Taking into account possible detuning Freq. detuning from f 0 Rev PA1 19
Impact of BW/Q: Reduction of the OOK Extinction Limited BW of ring reduces the ER of OOK modulation Impact on ER depends on the input ER (itself dependent on modulation) Q 20000 BW 3dB GHz Taking into account Bandwidth of ring and DATA RATE er in = A 2 /B 2 er out = A 2 /B 2 A B A B A B A A B B T b 2T b 3T b 4T b 5T b T b 2T b 3T b 4T b 5T b system NRZ OOK Distorted NRZ OOK Rev PA1 20
Impact of BW/Q: effect of other filters Not shown in db Attenuation of the Lorentzian tail for drop path at critical coupling (drop) Insertion Loss at the resonance (a function of Q factor) Extra attenuation by detuning from the resonance (a function of Q factor) Possible detuning Resonance frequency Attenuation for the through path at critical coupling (thru) Rev PA1 21
Total Penalty of a Single Add/Drop Ring Filter Increasing Q will increase both the IL penalty and Truncation Penalty R = µm @ critical coupling (any rate) R = µm Increase of DATA RATE decrease of Radius α loss = 2.8 db/cm Playing with the gap size Rev PA1 22
Impact of Q: Crosstalk in Ring Filters Optical Crosstalk as a noise mechanism Xtalk power (total) Target BER 0 th -order Approximation ignore the spectral bandwidth of modulation OOK light Effect of ER of OOK (sensitivity to noise) Assume all the optical power is at the carrier (center) wavelength Good approximation for LOW data rates and/or FLAT filters This method has been widely used NRZ optical power L. H. Duong et al., A case study of signal- to-noise ratio in ring-based optical networks-on-chip, Design & Test, IEEE, vol. 31, no. 5, pp. 555, 2014. L. H. Duong et al., Coherent crosstalk noise analyses in ring-based optical interconnects, in Proceedings of the 2015 Design, Automation & Test in Europe Conference & Exhibition, pp. 50150, EDA Consortium, 2015. J. Chan et al., Physical-layer modeling and system-level design of chip-scale photonic interconnection networks, Computer-Aided Design of Integrated Circuits and Systems, IEEE Transactions on, vol. 30, no., pp. 15071520, 2011. Rev PA1 23
Impact of Q: Crosstalk in Ring Filters 1 st -order Approximation do not ignore the spectral shape of the OOK modulation estimate the crosstalk power based on the Lorentzian shape of the ring based on the spectral shape of the NRZ OOK modulation based on the DATA RATE Fraction of crosstalk power from each NRZ channel data rate Rev PA1 24
Optimization results [Bahadori, et al. JLT, under revision] [Bahadori et al., Optical Interconnects, 2015] Rev PA1 25
Examples of Fabricated Rings P. Dong et al., Optics Express (2007) thru loss 0.4 db (measured), 0.003 db (model) drop loss 3.5 db (measured), 3.9 db (model) Widely cited R = 0 µm (very big ring) Q 19000 (measured), Designed for critical coupling But we could do better Reduce drop loss Pick the ideal BW for link Measurements do not match model (higher loss in ring in 2007) thru ER 18 db (measured), 12 db (model) Rev PA1 2
Another ring Analysed in Q. Li, PTL 27(18), 2015 t in = t out = 0.91 κ in = κ out = 0.44 Q = 1842 α = db/cm No critical coupling! Power penalties (measured): Thru: -0.1-0.8dB (channel dependent) Drop: -0.4-1.8dB ( Gbps signals) Rev PA1 27
Examples of fabricated ring: filter/modulator Q. Li et al., OFC (2014) Q. Xu et al., Opt. Express (2007) Modulator: R = 8 µm, Q = 4300 Modulator: R = 5 µm, Q = 20000! n eff Power of bit 1 Laser linewidth R g Rev PA1 28
Methodology - Abstraction of Physical Devices Input laser power, Number of wavelengths, Modulation rate Param Explore both device and link parameters to optimize bandwidth or energy efficiency Link Laser Modulator Detector P in Bandwidth Energy per bit Switch Demux Abstract Physical Models Link Model the flow and characteristics of optical signal along the link Rev PA1 29
Methodology - Energy Analysis (pj/bit) Energy Analysis Optical Loss Receiver Sensitivity Electrical Circuits Required Laser Power Transmitter Receiver FEC gain Energy Circle Thermal Tuning Serialization Modulators Thermal Tuning Amplifier Deserialization FEC encoder FEC decoder Rev PA1 30
Goal: conduct link wide optimizations E.g. extinction ratio (optical signal quality) vs. voltage (power consumption) For low number of wavelengths, largest resonance shift not required [1,2] [1] R. Wu, Rev et al. PA1 Compact modeling and system implications of microring modulators in nanophotonic interconnects, ACM SLIP 2015 31 [2] S. Rumley, et al. PhoenixSim: Crosslayer Design and Modeling of Silicon Photonic Interconnects, AISTECS workshop, 201
Conclusions Ring resonator based interconnects as very complex systems requiring fine tuning Point I was thinking to make at this workshop: mind the ring bandwidth! After making these slides: even more complex design space! For every ring: Input gap, output gap, radius, (doping) Find the right BW (depends on the architecture, bit-rate), align the wavelength, balance losses (also depends on the architecture), reach desired FSR Constant power penalty based approach questionable Can be very conservative, or very optimistic, depending on the context Silicon photonics still lacks maturity Well defined compact models, PDKs need to be defined Previously proposed designs should be re-assessed against these definitions Large scale modeling/design methodologies building on these PDKs to be developed The (current) impossibility to realize small rings is a big concern Rev PA1 32