Electromagnetic Modeling and Signal Integrity Simulation of Power/Ground Networks in High Speed Digital Packages and Printed Circuit Boards

Electromagnetic Modeling and Signal Integrity Simulation of Power/Ground Networks in High Speed Digital Packages and Printed Circuit Boards Frank Y. Yuan Viewlogic Systems Group, Inc. 385 Del Norte Road Camarillo, California 9300, USA fyuan@viewlogic.com. ABSTRACT The electromagnetic modeling and parameter extraction of digital packages and PCB boards for system signal integrity applications are presented. A systematic approach to analyze complex power/ ground structures and simulate their effects on digital systems is developed. First, an integral equation boundary element algorithm is applied to the electromagnetic modeling of the PCB structures. Then, equivalent circuits of the power/ground networks are extracted from the EM solution. In an integrated simulation scheme, the equivalent circuits are combined with signal nets, package models, device circuits, and other external circuitry for system level signal integrity analysis and simulation. This methodology has been implemented as software tools and applied to practical design problems. Effects related to power/ground networks, such as simultaneous switching noises, crosstalk, and ground discontinuity are analyzed for realistic designs. 2. INTRODUCTION With the ever increasing speed and density of digital integrated circuit systems, effects due to field interaction among IC chips, packages and PCB boards have become more and more the limiting factors in high speed system design [-3]. Signal integrity effects such as propagation delay, crosstalk, and simultaneous switching noises (SSN) require more accurate and efficient electromagnetic analysis and modeling beyond the traditional static method with a few lumped circuit elements. Particularly, the power/ground supply networks in digital circuits involve complex, large structures such as combinations of plane/partial planes and routed traces in multilayer environment, and are often tightly coupled with other parts of the circuit systems, such as signal networks and chip packages. Characterization of true distributed effect and frequency dependence is crucial. On the other hand, due to the overwhelming complexity of digital circuit board, full-wave rigorous electromagnetic field solutions, such as those used in microwave analysis [4], become impractical for digital application because of the extremely high computational requirement, and the difficulty of integrating with time-domain circuit type simulators and design environments used by digital designers. Therefore, EM modeling and simulation methodologies that can handle large complex digital designs with reasonable accuracy and efficiency, and can provide good macro models over extended frequency bands, are very much desired and thus present a challenge to us [5-8]. This paper presents a systematic approach which includes an efficient method for the electromagnetic modeling of packages and power/ground structures, and an integrated process for the analysis and simulation of system signal integrity and SSN effect in practical digital designs. The focus of this work is to develop a fast, flexible and accurate electromagnetic model with a clear circuit interpretation for the power/ground network, and to integrate it with active device models, transmission line networks, and chip package parasitics in a time domain co-simulation scheme. This methodology has been implemented and applied to practical design problems. It should be able to handle the complexity of real IC/ MCM/PCB designs within the practical computational constraints of an engineering workstation environment [7,8]. The paper is organized as follows: In Section 3, the electromagnetic modeling and analysis is presented. The general interconnects and power/ground structures are modeled as a multi-layered dielectric embedded with conductors. A mixed potential integral equation with layered Green s function is applied to the electromagnetic modeling of the structure, and is solved numerically using a boundary element approach. Section 4 describes the extraction of distributed equivalent circuits. A quasi-static approximation is introduced to simplify the complex frequency dependence of the power distribution system, while retaining the important AC characteristics. From the electromagnetic field solutions, equivalent circuits are constructed with frequency independent R, L, C circuit elements. This greatly reduces the required computational time, compared to a full wave solution, and makes a time-domain digital circuit simulation practical for real design applications. The distributed equivalent circuit models give accurate high frequency characteristics of the system up to a certain frequency limit that is well above most digital signal bandwidth, and can be used as macromodels for the power/ground supply networks. Section 5 describes the integrated simulation approach for system level signal integrity and SSN analysis. First, an efficient circuit solver is specifically developed for the equivalent circuit. A completely designed digital PCB system is partitioned into four major subsystems: chip devices such as drivers/receivers; chip packages; signal nets in the form of transmission lines; and power/ground supply nets in the form of equivalent circuits. Appropriate methods have been applied to model the different subsystems, and all the subsystems are linked together dynamically in a time-domain cosimulation. In Section 6, various examples are given to verify the approach against measurements and published results. Then, simultaneous switching noises and the effects of decoupling capacitors are analyzed in practical designs. 3. ELECTROMAGNETIC MODELING AND ANALYSIS 3. Electromagnetic Formulation Consider the general structure of integrated circuits and printed circuit boards as multi-layer dielectric planar substrate embedded -583-049-x-98/0006/$3.50 35 th Design Automation Conference Copyright 998 ACM DAC98-06/98 San Francisco, CA USA

with signal traces and power/ground planes as arbitrarily shaped thin conductors. The planes and signal traces are connected to each other and to external power supplies through vias or ground pins. Currents injected by active devices and/or power voltages applied to the conductors produce induced charges and currents on the conductor surfaces, which in turn generate induced electromagnetic fields. The electric field E(r) can be expressed in terms of a vector potential A(r) and a scalar potential Φ(r) as Er ( ) = jωar ( ) Φ( r) where A(r) and Φ(r) are related to the charge and current densities q(r) and J(r) by the following relation: Ar ( ) = µ dr G A r ) J( r ) Φ( r) = V -- d r G ε φ r )q( r ) V In the above equations, G A and G φ are the dyadic and scalar Green s functions for the vector and scalar potentials in the layered structure, respectively. The Green s functions have been derived analytically in forms of spectral integrals in general, and as infinite series for some special cases. Characterizing the lossy conductor by its surface impedance Z s, then the total electric field and surface current density J s (r) satisfy the following impedance boundary condition over the conductor surfaces: n E s () r = Z s n J s () r Also, the surface charge q s (r), surface current J s (r), and the external current flowing into the ports of the conductors, J i, satisfy the general continuity relation: J s ( r) + jωq s ( r) = J i ( r) Combining the boundary condition with continuity equation, we have the following set of mixed potential integral equations: Z s n J s () r + n jωµ dr'g A r' ) Jr' ( ) + n Φ() r = 0 (6) S J s ( r) + jωq s ( r) J i ( r) = 0 Φ( r) --- d r G (8) ε φ r )q s ( r ) = 0 S where S is the conductor surfaces and r S. This set of integral equations completely describes the potential, current, and charge distributions on the power and ground conductor planes. By solving these equations, the voltage and current distributions over the conducting planes can be obtained in relation to the injected currents or fixed port potentials for given excitations. Notice that the scalar potential is retained as an explicit variable in the above MPIE s. If we consider the scalar potential as a close approximation to voltage defined in a static sense, the above formulation enables us to obtain an explicit relation between port voltage and injected current, and thus a multi-port network representation or an equivalent circuit for the substrate system. 3.2 Boundary Element Method A boundary element method is applied to solve the integral () (2) (3) (4) (5) (7) equations. The total conductor surface, domain S, is divided into N elementary sub-domains (elements) with quadrilateral shapes. As an example, Figure shows the discretisation of a split MCM power plane where the 3.3 volt power net and the 5 volt power net complement each other. The current, charge and potential over S are expanded in terms of subsectional basis functions defined over each element, as N N N J s ( r) = I i T i q s ( r) = Q i ϕ i Φ s ( r) = V i φ i (9) i i i where I i, Q i and V i are the unknown expansion coefficients, and T i, ϕ i and φ i are basis functions for current, charge and potential, respectively. Note that current is a vector quantity, thus T i is a vector basis function with both x and y components. There are many possible choices for basis functions. The simplest is pulse function which will be used in the expansions of charge and potential distributions. To ensure the continuity of current, however, a bilinear expansion will be utilized for current expansion. After substitution of the above expansions into the integral equations, an inner product of a test function at each integral equation is taken over each subdomain (testing procedure or sampling). The choice of test functions is also a crucial matter. The simplest choice is a delta function, corresponding to the socalled point matching method. It is computationally fast and simple, but exhibits accuracy and stability problems. A more sophisticated choice utilizes the same basis functions as the test functions, called Galerkin s method, which leads to improved accuracy and stability at the expense of computational requirement. Both methods have been implemented in our simulation program and used when as appropriate. Following the above boundary element procedure, the integral equations become a set of linear equations (in matrix notation) as ( Z s + jωl)i PV = 0 (0) P T I + jωcv = J i () where: Z s, L, P, and C are matrices; vector V represents the potential (voltage) while vector I represents the surface current on all the conductor surface elements; the element of J i vector is the external current injected into each surface element, which except on terminal elements (ports) is zero. The elements of Z s, L, P, and C matrices in general involve two dimensional double surface integrals of Green s functions and test functions over the subdomains. Special techniques such as closed form formulas have been applied in the evaluation of those integrals. Depending on the specified boundary (terminal) condition, we can solve for either the unknown port voltages with specified port currents, or the unknown port currents with specified port voltages, along with the charge, surface current and potential all over the conductor planes. 4. QUASI-STATIC APPROXIMATION AND EQUIVALENT CIRCUIT EXTRACTION 4. Quasi-Static Approximation The above integral equation/boundary element approach is a rigorous full-wave analysis as it has been presented up to this point. This formulation of the system, as represented by the matrix equations, has complex frequency dependence. All the matrix elements and operations are frequency dependent and thus need to be recomputed and updated for every single frequency component for wideband or transient applications.

() we have the system solution as V i = ZJ i, where Z = jωc + P T ( Z s + jωl) P (7) Taking a Taylor expansion of (7) in frequency and keeping only the first and second order terms, we obtain the following Z = jω( P T L P) ( jω) 3 ( P T L P) CP ( T L P) (8) (a) (b) Figure. Split MCM power planes and their discretisation. (a) 3.3 volt VCC0 net, (b) 5.0 volt VCC net. They share a common ground plane with a dielectric layer of 0.5 mm thick. For a time-domain digital circuit simulation on the scale of MCM/ PCB substrates, the above full-wave solution requires a tremendous amount of computation. A simplification in the frequency relation is therefore very much desired. For typical MCM/PCB substrate environments, the following relation provides a guide to the applicability of quasi-static methods: h -- = hf εµ «(2) λ where h is the characteristic dimension of the sub-domain structure and f is the operating frequency. Thus for systems satisfying (2) it is reasonable to make the following quasi-static approximation. First, the higher order terms in the exponential retardation factor are neglected in both the vector potential Green s function and the scalar potential Green s function. For example, the free space Green s function G(r) = exp(-jkr)/4πεr is replaced by the static one as G(r) = /4πεr. Under this approximation, the matrix elements in L and C become real and frequency independent. The only frequency dependence is now the jω factor in the matrix equations which will be further simplified. Secondly for good conductors, we have Z s + jωl R s ( ω 0) Z s + jωl jωl ( ω» 0) (3) (4) With this assumption, the surface resistance Z s can be considered as the first order low frequency approximation for the total resistance, and the system becomes a pure resistive network at DC. Solving equations (0) and () under the above quasi-static approximation, we obtain the system solution in admittance form as J i = YV i, where J i and V i are the port currents and voltages, respectively, and Y is the system admittance matrix given by Y = jωc + P T ( Z s + jωl) P (5) Let L T = ( P T L P), and neglect surface loss, then the Y matrix has the following form as Y = ------ ( L (6) jω T ) + jωc where ( L T ) denotes a matrix with elements equal to the inverse of the corresponding elements of L T, not the inverse matrix of L T. Similarly, an impedance matrix formula can be obtained under the same quasi-static approximation. Again, from equations (0) and Let L R = ( P T L P), we finally have the simplified impedance matrix as Z = jωl R ( jω) 3 L R CL R (9) In the above formula, all major matrix operations are frequency independent. More sophisticated expansion and manipulation in the above procedure could include the surface resistance term, and lead to similar results with simple frequency dependence. 4.2 Equivalent Circuit Extraction From the impedance matrix or the admittance matrix, a distributed equivalent circuit representing the system can be constructed. Since a nodal formulation based on Kirchhoff s current law (KCL) is preferred over a mesh formulation, we consider the extraction of equivalent circuit from the admittance matrix Y as in Equation (6). In fact, the admittance matrix is the same Y matrix in a nodal formulation of the circuit in frequency domain, and can be used directly for frequency domain solutions [9]. To see more clearly the equivalent circuit in terms of its R, L and C elements, and to use it for more general circuit purposes, note the congruence between the admittance representation in Equation (6) and the frequency domain admittance expression for a parallel LC circuit branch, as given below in Equation (20). The equivalent circuit thus consists of branches between every pair of ports. Each branch has an inductance L in series with a resistance R, and then in parallel with a capacitance C. A four node equivalent circuit with common ground is shown in Figure 2. The branch admittance between nodes m and n is given by Ỹ mn = -------------------------------- + j ω C mn R mn + jωl mn (20) For circuit analysis, write KCL at each node. This leads to the same circuit equation as obtained by rewriting the admittance relation in terms of differential voltages between nodes M J m = Ỹ mm V m + Ỹ m ( V m V ) + = Ỹ mn V mn n = (2) The branch admittances Ỹ mn are then determined by comparison with matrix Y, as following Ỹ mn = Y mn ( m n) (22) M Ỹ mm = Y nm (23) n = Correspondingly, the L and C values are obtained from the matrix elements in L T and C as

L mn = L mn ( m n) (24) C mn = C mn ( m n) (25) while L mm = 0 (26) M C mm = C nm (27) n = Equation (26) follows from Kirchhoff s current law for the reference node. To a first order approximation, the resistance R takes the corresponding values of the DC resistance matrix obtained through (5) as P T ( Z s P). For a real design where every external connection, such as power/ground pin, is selected as a circuit node, then the equivalent circuit is a N node network where N is the number of power/ground connections. Since the equivalent circuit captures the wideband frequency behavior and distributed characteristics, it should correctly predict the wave effects such as propagation delay, resonance, and signal reflection on the power/ground plane structure. Numerical examples show that it indeed captures those effects correctly. However, it is fundamentally limited by the retardation effect, and does not include losses due to radiation. 5. CIRCUITS AND SYSTEM LEVEL SIMU- LATIONS 5. Circuit Solution The equivalent circuits can be utilized to simulate the response of complex interconnect and power/ground structures, or be used as subcircuits or macromodels in integrated system level simulation. An efficient circuit simulator specifically for the equivalent circuit is developed in our simulation program which provides both timedomain and frequency domain capability. The time-domain circuit solution algorithm is developed based on modified nodal formulation [9]. The linear equivalent circuit system is characterized by Gx + Cx' = w (28) In the above equation G consists of resistive and conductive elements while C consists of capacitive and inductive elements. Special formulation of the system equations eliminates the unnecessary internal inductance nodes and therefore greatly reduces the rank of the linear system. Both first order and second order integration methods are used in solving the linear circuit equation, providing good stability and accuracy with speed. Combined with uniform time step for the linear circuit portion, this approach gives us very efficient simulation time and thus enables us to analyze the large scale networks from realistic circuit boards and designs. General purpose circuit simulators such as SPICE can also be used for the simulation, but may not be very efficient. Frequency domain simulations are useful for gaining insight of high frequency characteristics of the complex systems. In practice, experimental measurements and characterization of high-speed/ high-frequency systems are mostly made in frequency domain in terms of S-parameters. It is thus used for verification of the simulation code s accuracy and limitation in comparison with experimental measurements and other benchmark results. Figure 2. A four node equivalent circuit with common ground. 5.2 Integrated System Simulation In our integrated simulation scheme, given a completely designed digital PCB board, the system can be partitioned into four major subsystems: chip devices such as drivers/receivers, chip packages, signal nets in form of transmission lines, and power supply nets in the form of planes/partial planes and routed traces. Proprietary behavioral device models, as well as IBIS or SPICE models can be used for IC chip devices or discrete components with various accuracy. Accurate device models are critical to correct simulation and many IC vendors provide models or data for their products in one or other format. Chip package modeling involves mostly parasitic extraction for parameters such as pin inductance and capacitance, and the package is modeled as a few circuit elements or a subcircuit connecting devices with signal traces and power/ ground nets. The signal nets are modeled as multiconductor transmission lines (microstrip or strip lines). Fast 2-D field solver is used to extract the transmission line parameters, and an accurate and efficient modal analysis is applied to the time-domain simulation of signal propagation which includes crosstalk between multiple lines. The subsystems are very different in their operation principles, their characterizations and behaviors, and the ways they are analyzed or simulated. Various and appropriate methods have been applied to different subsystems along with focus on various aspects of interests such as signal crosstalk or package extractions. All the four subsystems interact dynamically with each other in the system operation, and therefore are combined together for system level simulation in our analysis, as shown in Figure 3 below. Chip Packages IC Chip & Devices Power/Ground Planes Signal Network Figure 3. Partition and interaction of the four subsystems of a PCB circuit.

The details of the modeling and simulation methods in the other subsystems except power/ground networks are omitted here due to limited space, but will be presented separately in future publications. In the case of simultaneous switching noise, when one or more drivers on a single or multiple chips switches, large transient currents are drawn through package pins/connections from the power distribution network (planes) and thus create noises or voltage fluctuations in the power supply, which in turn affect the operation of the switching devices as well as other devices. The switching currents therefore act as the excitation sources to the distributed power/ground planes and the transient noises propagate and resonate in the plane structures. In our co-simulation, since every power/ground pin is a node in the equivalent circuit, at every time step the driver Vcc and Gnd currents are imposed upon the power/ground net as source to calculate the ground noise responses, and these noises are fed back to the device and signal simulation, allowing dynamic interaction among all the subsystems. 6. VERIFICATION AND APPLICATIONS The following examples first verify the approach against measurements and known results in both time and frequency domain analyses. Then applications on practical designs will be shown for the effect of simultaneous switching noises. 6. Comparison and Verification As the first example, we simulated the case of L-shaped microstrip patch from [4]. A three port equivalent circuit gives good low frequency results comparable with [4]. To better capture the high frequency characteristics, a 4-node equivalent circuit is obtained and simulated to calculate the resonant response for the input impedance at node A. The first two resonant modes are found at f 0 =.02 GHz and f =.65 GHz, respectively. Comparing with the full wave results f 0 = 0.998 GHz and f =.56 GHz given by [4], it demonstrates that the equivalent circuit gives good frequency response up to the first several resonant modes, which are well above the bandwidth of most PCB digital circuits. As the second example, the coupled microstrip line structure shown in Figure 4 was simulated. A 6-node equivalent circuit was extracted and used in the transient simulation of a pulse signal propagating along the lines. A voltage source with a 50 ohm output resistance was used as driver for the active line, while the other ends were matched with 50 ohm loads. A 5 volt pulse signal of 0.3 ns rise/fall time and.0 ns duration was applied at the near end input. Figure 5(a) shows near and far end responses on the active line, while Figure 5(b) shows the near and far end crosstalk on the passive line. Those results were compared with commercially available transmission line simulator and good agreements were obtained. It needs to be pointed out that for well known structure like microstrip line, more efficient and natural approaches exist and should be readily applied to similar problems. This example, however, demonstrates the correct prediction of propagation delay characteristics and coupling effects from the distributed equivalent circuit model extracted with a quasi-static approximation. 6mm 6mm 6mm ε r = 4.5 5mm Figure 4. Cross section of the coupled microstrip line in example 2. Voltage (volt) Voltage (volt) 3.0 2.0.0 0.0 0.2 0. 0.0-0. -0.2 "N_end.dat" "F_end.dat" 0.0 0.5.0.5 2.0 (a) Time (ns) 0.0 0.5.0.5 2.0 (b) Time (ns) Figure 5. (a) Near and far end voltage waveforms on the active line, (b) near and far end crosstalk waveforms on the passive line. 8 mm p p2 8 mm p3 p4 p5 probing pads "N_end.dat" "F_end.dat" 280 µm Figure 6. Test structure by HP Lab. The planes are made with 6 mω/sq. tungsten. The dielectric is alumina with ε of 9.6. In the third example, we compare our simulation results with the measurement data for the test plane structure given by K. Lee et al in [3], as shown in Figure 6. In our simulation, a 42-node equivalent circuit was extracted from the field solution, and used in frequency domain calculation of S-parameters. Figure 7 compares the simulated and measured results, where the measured S2 was given in [3]. It is evident that the agreement is quite good up to about 0 GHz, and the simulation captured the essential characteristics of the system. Towards higher frequency, however, the simulated result shifted away from the measurement in a systematic fashion. This behavior is within our expectation due to the limit of quasi-static approximation. But for most current digital applications, as we pointed out before, 0 GHz is a sufficiently

high estimate of the signal source bandwidth while the computational requirement is practical. A full-wave simulation, though possible under the current framework, would be too costly and unnecessary. To further study the transient characteristic of the system, and to see the effect of high frequency limit on high speed digital signal, time domain simulations using both the equivalent RLC circuit and 2-D FDTD are carried out on this test structure. A pulse signal of 5 volt magnitude, 0.2 ns rise/fall time, and.0 ns duration is applied at Port. All five ports are terminated with 50 ohm load. Transient waveforms are calculated. In the FDTD simulation, a grid size of mm by mm and a time step of 0 ps are used. Figure 8 shows the voltage waveforms at Port 2, where solid lines are the results obtained by equivalent RLC circuit while dashed lines are FDTD results. Good agreement again is evident and demonstrates the validity of our approaches. 6.2 System Level Simulation of SSN Power and ground noises due to the inductive effects in packages and PCB boards have become a great concern for practical digital system design. A most common practice in PCB design is the utilization of decoupling capacitors to reduce switching noises. However, due to the lack of analysis tools, the de-caps are used in a way of play it safe and put as much as you could if there is space in practical design. A major application for this work is to simulate the effect of de-caps and thus optimize the decoupling strategy which includes the placement, number, and value of decaps necessary for noise reduction against design margin. As a practical example of pre-layout evaluation, we simulated the SSN and decoupling effects of a 7x0 inch, six layer FR4 board with power and ground planes separated by 30 mil. A chip with sixteen CMOS drivers is placed on the board. The ground noises were simulated with different combination of drivers switching, and the effectiveness of decoupling capacitance were observed. As another practical application, a real customer design was used for post-layout system signal integrity simulation and evaluation of power/ground network. This four-layer board with twenty-six chips has two power/ground planes separated by 0 mil and has 55 Vcc pins and 80 Gnd pins. Limited by space, details are omitted here for these two examples. Refer to the presentation and [8] for more complete description of the designs and the simulation results. 7. CONCLUSION The modeling of power/ground networks and simulation method for system level signal integrity in high speed digital packages and MCM/PCB s are presented. This systematic approach enables us to analyze effects such as simultaneous switching noises, crosstalk, and the effectiveness of decoupling capacitors. The simulation results compare well with measurements and published results. 8. REFERENCES [] W. Becker et al, Power Distribution Modeling of High Performance First Level Computer Packages, IEEE 2nd Topical Meeting on Electrical Performance of Electronic Packaging, pp. 203-205, Monterey, CA, Oct. 993. [2] K. Lee and A. Barker, A Comparison of Power Supply Planes in Thick and Thin MCM s, IEEE 3rd Topical Meeting on Electrical Performance of Electronic Packaging, pp. 3-6, Monterey, CA, Nov. 994. [3] R. Senthinathan et al, Reference Plane Parasitics Modeling and Their Contribution to Power and Ground Path Effective Inductance as Seen by the Output Drivers, IEEE Trans. Microwave Theory Tech., vol. 42, pp. 765-773, Sept. 994. S2, S2 (db) 0-20 -40-60 -80 sim exp -00 0.0 2.0 4.0 6.0 8.0 0.0 Frequency (GHz) Figure 7. Comparison of simulated and measured S-parameters on the test plane structure. Figure 8. Comparison of transient waveforms at Port 2 on the test plane, obtained by FDTD and equivalent circuit simulations. [4] J. R. Mosig, Arbitrarily Shaped Microstrip Structures and Their Analysis with a Mixed Potential Integral Equation, IEEE Trans. Microwave Theory Tech., vol. 36, pp. 34-323, Feb. 988. [5] A. E. Rueli, Circuit Models for Three-dimensional Geometries Including Dielectrics, IEEE Trans. Microwave Theory Tech., vol. 44, pp. 263-267, Feb. 992. [6] G. Coen et al, Automatic Derivation of Equivalent Circuits for General Microstrip Interconnection Discontinuities, IEEE Trans. Microwave Theory Tech., vol. 44, pp. 00-06, July. 996. [7] F. Yuan et al, Analysis and Modeling of Power Distribution Networks and Plane Structures in Multichip Modules and PCB s, IEEE Electromagnetic Compatibility Symposium Proceedings, pp. 447-452, Atlanta, GA, Aug. 995. [8] F. Yuan, Analysis of Power/Ground Noises and Decoupling Capacitors in Printed Circuit Board Systems IEEE Electromagnetic Compatibility Symposium Proceedings, pp. 425-430, Austin, TX, Aug. 997. [9] J. Vlach and K. Singhal, Computer Methods for Circuit Analysis and Design, Van Nostrand Reinhold, New York, 994.