964 IEEE JOURNAL OF QUANTUM ELECTRONICS VOL QE-22, NO JUNE 1986 Architectures for Large NonblockingticalSpace Switches RONA Abstract-This paper introduces three architectures for optical space switches that are based on a multiplicity offiber interconnected optical components.Thearchitectureseliminate the need foropticalwaveguide crossovers and reduce the complexity required in the individual elements The architectures are strictly nonblocking and allow easy for control and routing Architecture type exhibits a low system attenuation and a high system signal-to-noise ratio for very large switch dimensions Architectures and are alternatives for realizing broadcast and point-to-point architectures SPANKE switch dimensions without complex integration onto individual LiNbO, substrates The three architectures have widely different attenuation and SNR characteristics, one of which exhibits excellent performance in both areas BASICOPTICALCOMPONENTS I :N passive splitters divide optical power evenly into N channels, while N : passive combiners combine the power of N inputs into a single output These passive devices INTRODUCTION could be realized with guided wave devices on LiNb03 EVERAL architectures have been proposed for optical[9] or by fiber couplers [lo] Passive splitters incur at least space switches fabricated on titanium-diffused lithium a 3k (dB) power loss from input to each output where k niobate (Ti:LiNb03) substrates [ 11-[4] These architecrepresents log2N Most present devices that perform the tures have been constructed using various optical switchpassivecombinersfunctionalsoincura 3k (dB)power ing elements including directional coupler switches [5], loss from any given input to the output totalinternalreflection(TIR)switches [6], andcrossing X N active splitters and N X active combiners are X switches [7] In addition, many of the classical switch- fabricated with X and X optical switch elements ingarchitecturesfound in theelectronicandcommunion Ti:LiNbO, These elements could be directional coucationdomainscouldbeimplemented withphotonic plers, X switches, or any other photonic switch element switching elements in the optical domain Such architec- Each X N splitter or N X combiner is constructed tures include the Clos, Benes, Banyan, omega, and shuffrom N - switch elements arranged in a k-stage binary fle networks to name a few tree configuration (Fig 1) To switch from the input port All of these architectures could be fabricated optically; to the required output port, k switches (one in each stage) however, they all encounter difficulties when expanded to need to be activated All of the elements in a given stage largeswitchdimensionswheretheybecomelimited by could possibly be electrically tied together so that only k systemattenuation,systemsignal-to-noise ratio (SNR), control leads and electronic drivers would be required inand LiNbO, real estate In addition, many of these archi- steadof N - A1 X 16 polarization-independent optecturesrequirethesignalpathstocrossthroughone ticalswitchwith15directionalcouplers in afour-stage anotherontheopticalsubstratebetweentheswitching binary tree structure has recently been demonstrated [ 111 elements.Thesepassiveintegratedopticalwaveguide crossovers appear feasible [SI; however, they contribute ARCHITECTURES additional attenuation and crosstalk problems for these arThe three optical switching architectures described here chitectures Optimal switching architectures in the elecare tronic domain where one attempts to minimize crosspoints constructed of active splitters and active combiners 2), are not the optimal switching architectures in the optical (type l ) , passive splitters and active combiners (type or active splitters and passive combiners (type 3) A fouth domain where one attempts to minimize attenuation and type constructed of passive splitters and passive combimaximize SNR ners is merely a novel method of implementing an N X N This paper proposes three space-division optical transmissive star.Itis not switchablewithoutanaddiswitching architectures based on passive splitters, passive tional stage of switches on each line in the center of the combiners,activesplitters,andactivecombiners All switch and will not be considered in this paper threearchitecturesuseamultiplicity ofinterconnected Architectures type consists of N X N active splitters optical components This eliminates the need for any integrated optical waveguide crossovers and allows for large on the left interconnected by fiber to N N X active combiners (Fig 2) to form a nonblocking optical interconnection network Types and replace the active splitters or Manuscript received October 11, 1985; revised February 4, 1986 the active combiners with their passive counterparts An The author is with AT&T Bell Laboratories, Naperville, IL 60566 N X N switch of type requires 2N(N - 1) optical switch IEEE Log Number 8608238 S 001 8-9197/86/0600-0964$01 00 O 1986 IEEE SPANKE:ARCHITECTURES 965 FOR OPTICALSPACESWITCHES I insertionlosses given by ZL forthethreearchitecturesarethen type 1: Z L = * k * L + *W(dB) types and 3: ZL = k (3 + L + E) + * (3) W (dB) (4) I Fig I X I N activesplitter using k stages of X opticalswitchelements in a binary tree structure where W represents the waveguide-to-fiber coupling loss which is roughly 1-2 dB at each interface The insertion losses for types and assume fiber devices for the passive components If guided wave LiNbO, devices are used forthepassivecomponents,types and wouldhave four waveguidehber transitions For comparison purposes, the worst case attenuation for [ ] thatusesswitchelements thecrossbararchitecture having an insertion loss of L is given by crossbar: ZL = (2N - 1) L +2 W(dB) (5) SNR CHARACTERISTICS Architecture type benefits from the selection of the desired signal in both the active splitter side and in the active combiner side Every optical input that is not the desired input for a given output represents a noise signal Fig N X N optical switch (type I ) for that output The double selection forces every possible source of noise to be attenuated by at least two optical elements, while and require N ( N - 1) switch elements extinction ratios before leaking into a desired signal This and N ( N - 1) 50/50 splitting(c0mbining) elements squared crosstalk isolation allows a large optical switch Because optical switch types and use an active split- to have a better signal-to-noise ratio than the individual switch elements that make up the switch ter, they function as a strictly nonblocking point-to-point switch No rearrangement of any optical signal paths is Allsignalandnoisepathsthroughtheswitchpass ever required Type2 uses passive splitters which provide through the same number of devices (2k) The insertion every input signal to every output combiner The active loss of these devices will appear in both the signal and combiner will then select the desired signal to leave the noise terms and will cancel out of the final SNR The inoutput.Architecturetype canthereforefunctionasa sertion losses due to the four fibedwaveguide transitions strictly nonblocking broadcast switch Any output can lis- will also appear in both the signal and noise terms and ten to any input, even if other outputs are currently liswill cancel out of the final SNR The power at the input tening to the same input of each X N active splitter is PIN.The power of each of the N signals leaving the active splitter will differ based ATTENUATIONCHARACTERISTICS on the number of switch elements passed with an extinction ratio X (power transmittance, i.e., -20 dB = 0.01) All three architecture types have insertion losses proportional to k instead of proportional to the switch dimen- The number of switch extinctions encountered will vary sion N The insertion loss in an active splitter or active from zero for the desired signal channel tok for the most combiner depends on the number of optical switch eleattenuated The number of channels with various extincments that the optical signal passes through and is given tion ratios will follow a binominal distribution (Table I) Given the fiber interconnect pattern between the splitby ting and combining' stages, it can be shown that at the LAs L A C = k L (dB) (1) inputofevery N X combiner, the N inputswillalso where L represents the insertion loss in a given switch exhibitthisbinomialdistributionofswitchextinctions element This L includes the straight waveguide attenuaInside the active combiner, the desired signal will pass tion (in dB/cm) due to scattering losses, the waveguide with no further extinction, while each of the noise terms bending losses, and the losses dueto incomplete coupling will encounter from to k additional switch extinctions that are associated with each switch element The worst case SNR will occur when thek highest power The insertion loss in a passive splitter and passive com- noisechannelsfromtheactivesplitter PIN X pass biner is given by throughtheactivecombinerwithonlyoneadditional switch extinction The binomial distribution of worst case Lps = Lpc = k * (3 + E ) (dB) (2) power levels at the N X active combiner output is given The dB figure represents the 50/50 power split and E in Table 11 represents the excess loss in each passive power split The The total noise power for architecture type is the sum - 966 IEEE JOURNAL OF QUANTUM ELECTRONICS, VOL QE-22, NO 6, JUNE 1986 TABLE I DISTRIBUTION OF POWER LEVELS LEAVIYG I I I Power(,,) Splitter X N ACTIVESPLITTER A Number of Channels General 1 (desired channel) N=64 - I 15 20 15 k.(k-l) .:( k-4 k TYPE of all N - unwanted channels that find their way to the worst case output and is given by c Ignoring X4 and higher order crosstalk terms, the case SNR for type can be approximated by SNR,,,, = * X,, - 10 * logk (dB) worst (7) I I R \ 16 32 I I I I I G4 128 ?5G 512 1024 21318 W I T C H DIMENSION N (b) Fig System attenuation (a) and system signal-to-noise ratio (b) versus switch dimension for architecture types1 , 2, and conventional crossbar switches Type 2, being a broadcast architecture, does not select the desiredsignal in the splitter stage and does not benefit from the double crosstalk isolation in the SNR character- Therefore, the noise power and SNR for architecture type are the same as for type istics The opticalpowerenteringtheactivecombiner For comparison purposes, the worst case SNR of the stage is the same for the signal channel and all noise chancrossbar [3] implemented with optical switch elements that nel and is represented by Pcomb have an extinction ratio of X,, is given by P c o m b = PIN - k ( + E ) (dB) (8) SNRcrossbar = x,, - 10 lOg(N - 1) (dB) (11) The noise power for type is the sum of the N - I unwanted signal powers entering an active combiner and is SWITCHDIMENSION LIMITS also binomial in nature Fig 3(a) shows the system insertion loss versus switch dimension N [ ( ) and (4)] for the three architecture types presented here The values for the switch element insertion losses, excess losses, and fiberlwaveguide coupling losses are assumed to be E = L = dB and W = dB A first-order approximation for the SNR yields For this comparison, the maximum attenuation allowed SNR,,,, = X,, - 10 1Ogk (dB) (10) from system input to output without amplification or reArchitecture type experiences the same binomial distri- generation is assumed to be 30 dB Fig 3(b) shows the [ ( )and (lo)] for bution of extinction ratios leaving the active splitter stage system SNR versus switch dimension the three architecture types with x,, = 20 dB The reas shown in Table I Given the fiber interconnect pattern quired SNR is assumed to be greater than 11 dB to achieve between stages, this binomial distribution of couprer exbiterrorrate.Thefiguresalsoshowinsertion loss tinctions is also present at the inputs to everyN : passive a combiner.Thepassivecombinercombinesthe signal and SNR characteristics [(5) and (ll)] for conventional crossbar switches channel with these N - channels of noise with an inWith E = L = dB, W = dB, and x& = -20 dB, herent k ( E)(dB) attenuation on allchannels + SPANKE:ARCHITECTURESFOROPTICALSPACESWITCHES a maximum switch dimension of 32 X 32 should be possible for architecture types and 3, yielding a system attenuation of 29 dB and a 13 dB SNR With the same device characteristics, a switch dimension of 1024 X 1024 is theoretically possible with type using ten-stage active splitters and ten-stage active combiners This yields a 28 dB system attenuation with a 30 dB syskm SNR Even with very large switch dimensions, SNR is not a limiting factor, and the system SNR for type1 is considerably better than the extinction ratios of the optical switch elements themselves.Inthisexample,webasedthemaximum switch dimension for architecture only on the insertion loss, but in reality, fiber interconnection now becomes the limiting factor Largerswitchdimensionscouldpossiblybe realized without complex LiNb03 integration and unreasonable fiber I/O by separating the active splitter and active combiner functions into two stages each This increases the overall optical chip count and system attenuation, thereby reducingthemaximumswitchdimension,but it allows for simpler components and easier system interconnection CONCLUSION Three space division optical switch architectures have been presented based on active and passive splitting and combining elements Type is a point-to-point switch and exhibits excellent SNR and attenuation characteristics for very large switch matrices Type is a broadcast switch exhibiting moderately good SNR and attenuation characteristics up to a 32 X 32 switch Type isapoint-topoint switch exhibiting the same system characteristics as type and is useful in matrices up to 32 X 32 The type and type architectures appear to be well suitedforlargepoint-to-pointandbroadcastoptical switches These architectures offer larger switch dimensions and improved attenuation and signal-to-noise performance over other optical switching architectures proposed 967 REFERENCES [I] M Kondo et al., “Integrated optical switch matrix for single-mode fiber networks,” IEEE J Quantum Electron., vol QE-18, pp 17591765, Oct 1982 [2] R A Spanke and V E Benes, “An N-stage planar optical permutation network,” to be published [3] H S Hinton, “ A nonblocking optical interconnection network using directionalcouplers,” in Proc GLOBECOM, 1984, pp.26.5.126.5.5 [4] L McCaughan and G A Bogert, “4 X strictly nonblocking integrated Ti:LiNbO, switch architecture,” in Proc OFUOFS ’85, San Diego, CA, Feb 1985, p 76 [5] R C Alferness, R V Schmidt and E H Turner, “Characteristics of Ti-diffusedlithiumniobateopticaldirectional couplers,” Appl Opt., vol 18, pp 4012-4016, Dec 1, 1979 [6] C S Tsai,B Kim,and F R El-Akkari,“Opticalchannelwaveguideswitch and couplerusingtotalinternalreflection,” IEEE J Quantum Electron., vol QE-14, pp 513-517, July 1978 [7] A Neyerand W Mevenkamp, “Single-mode electrooptic X-switch for integratedopticswitchingnetworks.” in Proc.IEE2ndEuropeat? Con Integrated Opt., no 227, Oct 17-18, 1983, pp 136-139 [SI E E.Bergmann, L McCaughan,and J E Watson,“Coupling of intersecting Ti:LiNbOl diffusedwaveguides,” Appl Opt., vol 23, pp.3000-3003,Sept 1, 1984 [9] T Findakly and B V Chen, “Single-mode integrated optical X N starcouplers,” in Tech Dig Topical Meet Opt FiberCommun., 1983, paper ML-2 1101 S K Sheem and T G Giallorenzi, “Single-mode fiber-optical power Opt Lett., vol 4, p 29, divider: Encapsulated etching technique,” 1979 1111 J E Watson, “Polarization independent X 16 optical switch using Ti:LiNb03 waveguides,” in Proc OFCIOFS ’85, paper WK-4, Feb 1985,p.110 networks Ron A Spanke was born in Tulsa, OK on June 27, 1958 He received the B.S degree in mechanical engineering in 1980 and the M S degree in electricalengineeringin1982,bothfromOklahoma State University, Stillwater He joined the staff of AT&T Bell Laboratories Naperville, IL, in 1982 He is currently pursuing the Ph.D.degree inelectricalengineeringat Northwestern University, Evanston, IL, where he is investigating directional coupler based optical interconnection .. .SPANKE: ARCHITECTURES 965 FOR OPTICALSPACESWITCHES I insertionlosses given by ZL forthethreearchitecturesarethen... an inWith E = L = dB, W = dB, and x& = -20 dB, herent k ( E)(dB) attenuation on allchannels + SPANKE: ARCHITECTURESFOROPTICALSPACESWITCHES a maximum switch dimension of 32 X 32 should be possible... single-mode fiber networks,” IEEE J Quantum Electron., vol QE-18, pp 17591765, Oct 1982 [2] R A Spanke and V E Benes, “An N-stage planar optical permutation network,” to be published [3] H S