1470

IEEE JOURNAL OF SOLID-STATE CIRCUITS, VOL. 33, NO. 10, OCTOBER 1998

Analysis, Design, and Optimization of Spiral Inductors and Transformers for Si RF ICs
Ali M. Niknejad, Student Member, IEEE, and Robert G. Meyer, Fellow, IEEE
AbstractSilicon integrated circuit spiral inductors and transformers are analyzed using electromagnetic analysis. With appropriate approximations, the calculations are reduced to electrostatic and magnetostatic calculations. The important effects of substrate loss are included in the analysis. Classic circuit analysis and network analysis techniques are used to derive two-port parameters from the circuits. From two-port measurements, loworder, frequency-independent lumped circuits are used to model the physical behavior over a broad-frequency range. The analysis is applied to traditional square and polygon inductors and transformer structures as well as to multilayer metal structures and coupled inductors. A custom computer-aided-design tool called ASITIC is described, which is used for the analysis, design, and optimization of these structures. Measurements taken over a frequency range from 100 MHz to 5 GHz show good agreement with theory. Index Terms Monolithic inductors and transformers, optimization of Si inductors and transformers, spiral inductors and transformers.

I. INTRODUCTION ILICON integrated circuits (ICs) are nding wide application in the gigahertz frequency range. Modern bipolar (Bi), complementary metaloxidesemiconductor (CMOS), transistors, allowing and BiCMOS processes provide high Si radio-frequency (RF) ICs to compete with GaAs ICs in the important low-gigahertz frequency ranges. However, reactive the lossy Si substrate makes the design of high components difcult. Despite this difculty, the low cost of Si IC fabrication over GaAs IC fabrication and the potential for integration with baseband circuits makes Si the process of choice in many RF IC applications. The demands placed on portable wireless communication equipment include low cost, low supply voltage, low power dissipation, low noise, high frequency of operation, and low distortion. These design requirements cannot be met satisfactorily in many cases without the use of RF inductors. Hence, there is a great incentive to design, optimize, and model spiral inductors fabricated on Si substrates. This topic is addressed in this paper. Since the introduction of Si spiral inductors [1], many authors have reported higher performance inductors on Si substrates, primarily utilizing advances in processing technology. This has included higher conductivity metal layers to reduce
Manuscript received May 27, 1997; revised November 17, 1997. This work was supported by the U.S. Army Research Ofce under Grant DAAH04-93G-0200. The authors are with the Department of Electrical Engineering and Computer Science, University of California, Berkeley, CA 94720-1772 USA. Publisher Item Identier S 0018-9200(98)06998-4.

the loss resistance of the inductor [2], use of multimetal layers to increase the effective thickness of the spiral inductor and thereby reduce loss, the connection of multimetal layer spirals in series to reduce the area of the inductors [3], [4], low-loss substrates to reduce losses in the substrate at high frequency [5], and thick oxide or oating inductors to isolate the inductor from the lossy substrate [6]. Little, though, has been written on the analysis and optimization of these structures. In [7], an analysis approach is presented where an equivalent circuit for each segment of the spiral is calculated and the inductor is considered as an interconnection of such segments. The approach is limited, though, as many important effects are not included. For instance, nonuniform current distribution due to skin and proximity effects within each segment is not considered. In addition, the impedance to substrate is calculated using a two-dimensional approach, making it difcult to apply to arbitrary structures or to coupled inductors. Most past researchers have used measurement results on previously built inductors to construct models. While this technique is most practical, it does not allow the possibility of optimization, nor does it allow the circuit designer freedom to choose parameters such as inductance, resistance, capacitance, and . Alternatively, researchers have used commercial threedimensional electromagnetic simulators [8], [9] to design and analyze inductors and transformers. While this approach is accurate, it can be computationally very expensive and time consuming. This prevents the designer from performing optimization. In this paper, we present an accurate and computationally efcient approach to overcome some of these difculties. II. EQUIVALENT CIRCUIT FORMULATION A. Electromagnetic Formulation Consider a typical spiral inductor or transformer. It consists of series and parallel interconnection of metal segments. Applying Maxwells equations to the conductive portions of such a structure, we obtain [10] (1) is the applied eld, is the current density, is the where is the magnetic vector conductivity of the metal segment, potential, and is the electric scalar potential. The rst term of the above equation arises from the applied eld, the second term represents the internal impedance of the segments, the

00189200/98$10.00 1998 IEEE

NIKNEJAD AND MEYER: SPIRAL INDUCTORS AND TRANSFORMERS

1471

Fig. 1. Equivalent circuit model of spiral segments.

third term represents external inductance, and the fourth term is the capacitive term. The magnetic and electric potentials are integrals over the charges and currents in the circuit. In the time-periodic case, we have (2) (3) The above retarded potentials can be simplied to static potentials since at frequencies of interest, the exponential term is nearly unity. As they stand, though, the above equations are impractical since they involve integrating over all charges and currents, including those that ow through the substrate. If, term involving the free-space however, we replace the Green function with the appropriate Green functions over a multilayer substrate, the volume of integration reduces to metal segments containing the charge and currents. In [11], we derive the electrostatic Green function over a multilayer conductive substrate that can be used in (2).1 By expanding and in (1) using (2) and (3) and by considering the currents and charges in each segment separately, we obtain a system of linear equations [12]. Each equation has terms representing the internal impedance of each segment, as well as the mutual magnetic and electrical coupling between each segment and every other segment, and nally terms representing the magnetic and electric coupling to the substrate. B. Circuit Equation Formulation From a circuit point of view, (1) can be interpreted as shown in Fig. 1. To derive the two-port parameters of such a structure, we must solve the circuit equations by selecting
1 The magnetostatic Green function can be derived in an analogous way. The computation of the inductance matrix over a lossy conductive substrate is the subject of another paper.

an appropriate number of independent equations, using, for instance, nodal or loop analysis. Let us consider the topology of the network in relation to the spiral. For a typical segments. Each sided spiral of turns, there are segment has one branch associated with the series inductance and resistance and one branch associated with the substrate impedance. Assuming a dense capacitive coupling impedance branches of lossy matrix, this adds an additional capacitors from segment to segment. Hence, the total number The total number of nodes of branches in the circuit is simply one plus the number of segments, Consequently, the total number of fundamental loops or in the circuit equals Although loop equations are a more natural formulation, due to the mutual inductive nature of the circuit, the number of required equations is one order of magnitude larger than the number of node equations. Hence, nodal analysis is the method of choice. Writing the nodal equations at node results in (4) is the voltage at node is the In the above equation, is the lossy capacitive current owing into segment and The above equations impedance coupling segment and involve the voltages and currents in the segments, similar to modied nodal analysis. The voltage across each segment is related to the currents owing in the segments by (5) is an entry of the symmetric inductance matrix where The diagonal terms represent the internal impedance of each segment, whereas the off-diagonal terms represent the magnetic coupling between segments. The above system of equations yields an invertible complex frequency-dependent equations in (4) and (5), the matrix. Since there are system can be solved directly using Gaussian elimination with

1472

IEEE JOURNAL OF SOLID-STATE CIRCUITS, VOL. 33, NO. 10, OCTOBER 1998

pivoting or factorization, in time However, the matrix naturally partitions into four square matrices as (6) is a matrix with elements (7) is an upper triangular band matrix with its diagonal entries as 1 and with superdiagonal entries as 1. Rewriting (6), we have (8) (9) is the inverse of the inductance matrix. where is known a priori, and no explicit As we will see later, matrix inversion needs to be performed. In addition, due to the the matrix products involving and simple structure of Hence, in forming its transpose can be done in order (8) and (9), no explicit matrix products need to be formed. decomposition of Thus, we can solve (8) by using an and then solve (9) directly for . Therefore, decomposition of the solution of (6) is dominated by one matrix, which can be performed in time This an is more than eight times faster than solving (6) directly. C. Inductance Matrix must be computed to solve The inductance matrix (6). At low frequency, this matrix may be computed easily since the self- and mutual inductance of each segment can be approximated in closed form by using the geometric mean distance (GMD) approximation. Grover [13] provides formulas for such calculations for fairly arbitrary congurations. In [14], an explicit formula appears for the GMD between two nite thickness parallel rectangular cross sections. In a square spiral, all segments are parallel and these formulas sufce. However, for polygon spirals or for arbitrary structures, we need the mutual inductance between nonparallel segments situated in general in three-dimensional space. While the exact formula for laments placed in these congurations is known in closed form, the GMD concept is difcult to generalize in these cases, as the integrals are not easily evaluated. In addition, at high frequency, these approximations fail altogether as the current distributions become nonuniform and the GMD approximation fails. The nonuniformity in the current distribution is due not only to the skin effect but also to proximity effects of neighboring current segments. Hence, although many past authors [7] have analyzed and developed closed-form equations for current constriction in an isolated conductor, these results are not directly applicable to the spiral inductor due to proximity effects. In [14], a nice technique is developed to handle this problem without abandoning the closed-form equations available. The current distribution in the conductor can be approximated in a step-wise fashion by breaking up each segment into

subsegments of uniform current distribution. Although this technique is general and accurate, it is computationally expensive. Consider a typical ten-sided spiral with ten turns as an example. Such a spiral has 100 segments, and if we were to divide each segment further into ve parallel segments, this would involve 500 segments and a corresponding 500 500 matrix inversion. To alleviate the demand on computational resources, we will assume that the segment current distribution is only inuenced by a small set of neighboring segments. For a typical planar spiral, this might involve only two neighboring segments. We rederive the results presented in [14] with this approxisegment mation in mind. For a set of segments, let the denote represent the th subsegment of segment and let be the the number of such subsegments in segment Let segment and let be the total current current in the in segment With this notation, the voltage across the segment is given by (10) is the resistance of the In the above equation, conductor, is the Kroniker delta function, and is the and Let mutual inductance between subsegments denote the set of segments in the neighborhood of segment or the set of segments that inuence the current distribution of segment [15]. Equation (10) may be broken up as follows: (11) Inverting the rst term corresponding to the local voltage , we obtain at segment (12) The total current in segment is given by (13) If we make the reasonable assumption that the voltage is independent of the index across each subsegment (as was done in [14]), then (13) becomes (14) Equation (11) may be rewritten as (15) Repeating for all one generates the desired matrix equation. The above technique involves one matrix inversion going from (11) to (12) and another matrix inversion of reduced order going from (14) to (15). The latter matrix inversion can be neglected since the matrix is small. Assuming

NIKNEJAD AND MEYER: SPIRAL INDUCTORS AND TRANSFORMERS

1473

that matrix inversion is the most computationally intensive whereas the original operationand that it is done where is the technique presented in [14] is total number of segments and is the average number of where is subsegmentsthis technique is one plus the average number of neighbors for each segment. faster and demands Hence, this technique is less memory to store the matrix. For the example given earlier, a ten-sided, ten-turn spiral with two neighbors and ve subsegments per segment, this means that this technique is 370 times faster and more than 1000 times less memory intensive. D. Capacitance Matrix Techniques for the calculation of the capacitance matrix are well developed and have wide applications. However, most approaches consider the capacitance matrix of conductors in free space, or over an ideal ground plane. In [11], we present an efcient technique that is directly applicable to the case of a conductive substrate. With this approach, the true three-dimensional capacitance matrix of the device segments can be extracted and used in (6). Various effects can be included in the capacitance matrix calculation, such as the presence of substrate taps and substrate shields [16]. The important effect of substrate coupling between structures residing on the same substrate can also be included. For more details, refer to [11]. III. ASITIC: A COMPUTER-AIDED-DESIGN TOOL FOR INDUCTOR/TRANSFORMER DESIGN, ANALYSIS, AND OPTIMIZATION Much is known about the optimization of the technological parameters of a process for optimal inductor performance [7]. Thicker or higher conductivity metal improves the quality at low frequencies, whereas a higher resistivity factor substrate and thicker oxide help to isolate the device from the substrate at high frequencies. The optimization of the geometry of inductors and transformers, though, is more difcult. Even for simple structures, such as square spirals, there are several parameters to optimize, including the area of the spiral, the metal width and spacing, and the number of turns. Not much can be said in general since the optimal geometry depends on the frequency of operation. For instance, at low frequency, one usually uses the minimum spacing available to maximize magnetic coupling, but at high frequency, proximity effects and magnetic coupling favor a larger value of spacing [8]. This is illustrated in Fig. 2, where we plot the resistance of a typical spiral using two values of spacing. Similar considerations apply to the area of the spiral and the number of turns. At lower frequencies, large areas allow wider metal widths to meet a given inductance at lower values of series resistance and therefore higher But at high frequencies, the is dominated by the substrate and smaller areas are favorable. This is doubly true for highly conductive substrates that suffer from eddy-current losses at high frequency [8]. With the introduction of multimetal structures, such as shunt and series connected spirals, the situation is even more
Fig. 2. Normalized spiral resistance as a function of frequency. TABLE I APPROXIMATE PROCESS PARAMETERS

complicated. Thus, the optimization of such structures must be done carefully on a case-by-case basis. To aid this process, a custom tool has been developed. The techniques presented in this paper have been collected into Analysis and Simulation of Inductors and Transformers for ICs (ASITIC),2 a userfriendly computer-aided-design (CAD) tool designed to aid the RF circuit designer in the designing, optimizing, and modeling of the spiral inductor and transformers. The tool is exible, allowing the user to trade off between speed and accuracy. For example, ASITIC can be used to quickly search the parameters space of an inductor optimization problem. By working with the entire metal layout of the circuit at hand, ASITIC allows general magnetic and substrate coupling to be analyzed between different parts of the circuit. Parasitic metal extraction may also be performed, including the effects of the substrate, and the resulting SPICE le may be included in more accurate simulations. The nal design of a spiral may be exported for nal layout; this feature is convenient in the layout of complex geometry spirals (nonsquare spirals), since their hand layout is time consuming, especially when changes are made to the geometry of the spiral. IV. MEASUREMENT RESULTS A. Description of Process/Measurement Setup All spiral inductors were fabricated in Philips Semiconductors Qubic2 BiCMOS process. Approximate process parameters are given in Table I. Using on-chip cascade probes, the average two-port s-parameters of the spirals are measured from 100 MHz to 5 GHz. Since the measurement setup holds wafers in place using vacuum suction, the back plane of the substrate is effectively oating. The substrate surface is grounded by placing substrate taps near the spirals.
2 See

http://www.eecs.berkeley.edu/ niknejad.

1474

IEEE JOURNAL OF SOLID-STATE CIRCUITS, VOL. 33, NO. 10, OCTOBER 1998

Fig. 4. Layout and equivalent circuit models of spiral and pads used for calibration procedure.

Fig. 3. Layout of spiral and probe pads for measurement.

Pad capacitance is deembedded by subtracting out the opencircuit structure -parameters from the spiral -parameters. The calibration procedure we use is the standard approach and is reported widely. But there is an intrinsic problem with this calibration approach that can render the measurement results inaccurate. Fig. 3 shows the layout of a typical spiral setup for measurement. To deembed the large substrate capacitance and resistance of the pad structure, the open-pad two-port parameters are also measured, as shown in Fig. 4. In the same gures, we show a schematic of the pad parasitics along with the spiral equivalent circuit. Clearly, the open-circuit pad structure does not contain the parasitic coupling between the spiral structure and the pads. In addition, the pad-to-pad coupling is different in the two cases shown, since the spiral itself tends to shield the pads from one another. Hence, the pad parasitics are not correctly deembedded from the measurement results. To show this effect, we simulate a sample inductor with and without pads. Next, we deembed the pad parasitics from the simulation results with pads (just as one would in measurement), and the resulting extracted series inductance, resistance, factor are shown in Fig. 5(a) and (b). The inductance and value is in error but in both cases increases with frequency; the series resistance values, though, deviate dramatically as a function of frequency. Notice, too, that the trend comes out incorrect, increasing as a function of frequency instead

of decreasing. Also shown is the resulting error in the factor. In addition, this error can change the location of the peak -factor, resulting in suboptimal designs. Not shown are the equivalent extracted substrate impedance values. The measurement errors can produce dramatic differences here too, changing the substrate resistance by a factor of two. The substrate capacitance can actually change sign due to resonance phenomena. In the following sections, in order to match simulation results to measurements, we simulated our structures under the same conditions as they were measured. To solve these problems, we can of course always minimize the parasitic coupling between the device under test and the pads by physically moving the structure away from the pads and by placing isolating substrate taps between the structures. Nevertheless, this requires large areas to be consumed on the chip. While this calibration problem occurs for a BiCMOS substrate, simulation also indicates some problems on an epitaxial CMOS substrate. The magnitude of the error, though, is much smaller. For instance, the series and change much less, and only the substrate impedance changes drastically due to the calibration step. B. Extraction of Circuit Parameters from Measurement Data At each frequency of interest, the equivalent pi-circuit of Fig. 6(a) can be extracted from -parameters. By taking equivalent inductance and real and imaginary parts of resistance are extracted. Similarly, input and output shunt capacitance and resistance to substrate may be extracted by and taking real and imaginary parts of Since the circuit is passive, it is also reciprocal, so that This extraction procedure is widely used and is simple to implement since the electrical parameters of the spiral are uniquely determined at each frequency. The drawback, though, is that the extracted inductance is not physical and contains the effects of the capacitive coupling, which tends to boost the inductance as a function of frequency, similar to a parallel tank. The extraction technique presented in [17] is superior in this regard. In this paper, we use the standard extraction procedure to compare our simulation results to measurements. However, for our nal device models, we use frequency-independent circuits such as shown in Fig. 6(b), which contain parasitic elements so that our extraction of inductance is more physical. On the other hand, there is no unique way to assign the parameters

NIKNEJAD AND MEYER: SPIRAL INDUCTORS AND TRANSFORMERS

1475

(a)

(a)

(b) Fig. 6. (a) Traditional spiral inductor model. (b) Modied spiral inductor model.

Equally applicable, one may use the rate of change of phase at resonance (18) The above equations are derived using second-order resonant circuits. For higher order circuits, perhaps the most general denition is based on ratio of energy stored in the circuit to energy dissipated per cycle, or (19) values of our models since we use optimization to t a loworder frequency-independent model to a distributed structure over a broad frequency range. The traditional approach of extraction involves computing (16) (20) The above denition has the awkward property that the is zero at self-resonance. Since inductors are usually operated far from self-resonance, this does not present too many problems. But in some applications, the inductor is used as a resonant tank close to self-resonance. In such cases, it is more appropriate to dene the using a 3-dB bandwidth denition (17) This capacitance will resonate the device at the frequency By examining the rate of change of phase, one of interest can nd the equivalent (21) of To illustrate this approach, in Fig. 7 we plot the a typical spiral inductor (square spiral with nine turns of should be application The best approach to dening dependent. Our approach to extraction is based on (18). At each frequency of interest, an ideal capacitor is inserted in shunt with the inductor with admittance equal to the The resulting admittance becomes imaginary part of

(b) Fig. 5. (a) Extracted inductance and resistance of spiral inductor with and without calibration procedure. (b) Extracted quality factor with and without calibration procedure.

1476

IEEE JOURNAL OF SOLID-STATE CIRCUITS, VOL. 33, NO. 10, OCTOBER 1998

Fig. 7. Extracted quality factor (21).

Q of a nine-turn spiral by using (16) and

dimensions given in the next section) using the denition present in (21). In the same plot, we show the traditional using (16). As we approach self-resonance, the denition of two denitions deviate greatly. Notice also that the peak value of occurs at different frequencies for the two denitions. In the following sections on measurement results, the data are measured at room temperature. It is important to calculate at the frequency and temperature of interest [18]. At low frequencies, the temperature coefcient (TC) will depend on will change the metal TC, whereas at high frequency, the due to the TC of the substrate resistance. C. Square Spiral Inductors Since square spirals are the most common inductors in Si RF ICs, we begin by comparing our simulation results of the previous section with measurement results. Many square spirals were fabricated, and measurement results were compared to simulation. The spirals have the following geometry: m, constant width m, constant inner area m, with turns . A chip and constant spacing layout of the eight-turn spiral appears in Fig. 8. For example, Fig. 9 shows the measured and simulated -parameters for the eight-turn spiral. As can be seen from the gure, simulation -parameters match the measured -parameters well. Fig. 10(a) plots the extracted series inductance from both measurement data and simulation data for the ve- and eightturn spirals. Good agreement is found for all spirals. For all spirals, the inductance is an increasing function of frequency. In reality, the physical inductance decreases as a function of frequency due to the current crowding at the edges of the conductors, which leads to a decrease in internal inductance.3 The extracted increase in is due primarily to the coupling as a function of capacitance, which boosts the effective frequency. For the smaller spirals, the same coupling capacitance tends to increase as well, but the effect of the shunt For large substrate resistance is to lower the effective series
3 For a highly conductive substrate, the eddy currents generated in the bulk substrate also lead to a decrease in inductance as a function of frequency, as these currents partially cancel the magnetic eld generated by the device.

Fig. 9. Measured and simulated -parameters of the eight-turn square spiral.

spirals in which substrate injection occurs at a lower frequency, is a decreasing function of frequency. For the behavior of smaller spirals, the substrate injection is minimal due to a small substrate capacitance, and the effect of the coupling capacitance boosts as a function of frequency. Fig. 10(b) shows the -factor for the ve- and eightturn spirals. Here, we used the phase denition of (21) to Good agreement is found between measurements compute is an and simulations. For the smaller ve-turn spiral, the almost linear function of frequency, demonstrating that the is substrate has negligible effects on the spiral, and the The larger eight-turn spiral, approximately given by though, shows more complicated behavior due to the substrate. At low frequencies when the substrate impedance is large, simple linear behavior is observed. At high frequencies, when the substrate impedance is smaller than the inductive/resistive impedance of the spiral, the substrate loss dominates and the

NIKNEJAD AND MEYER: SPIRAL INDUCTORS AND TRANSFORMERS

1477

TABLE II SIMULATED AND MEASURED MODELS OF SQUARE SPIRALS

TABLE III POLYGON SPIRAL GEOMETRY

(a)

(b) Fig. 10. (a) Extracted inductance of square spirals from measurements and simulations. (b) Extracted -factor of square spirals from measurements and simulations.

Fig. 11. Layout of polygon spiral inductor L4.

D. Polygon Spiral Inductors is a decreasing function of frequency. In the frequency range of interest, the is a combination of the above-mentioned effects peaks at some frequency Notice that our and the analysis predicts this frequency well. This is important since at the frequency we can design inductors that have peak of interest. We model the inductors using the equivalent circuits of Fig. 6(b). Table II compares our results, including the total error involved in the modeling. The error is derived by computing the relative error between the two-port measured sparameters and the model two-port -parameters. The error is cumulative over 16 points in the frequency range. Two models were generated, one based on the actual measurement results and the other based solely on simulation results (the simulation model is not a curve-tted model of the measurements but rather a curve t to the simulations). Since polygon spirals with more than four sides have higher than square spirals (for the same area), it is advantageous to use these structures. Many spirals are measured and compared with simulation. The geometry of the fabricated spirals is shown in Table III. The chip layout of a sample spiral inductor L4 appears in Fig. 11. -parameters are plotted for the L4 spiral in Fig. 12, and again good agreement is found between simulation and measurement. Extracted series inductance and for the L4 and L11 spirals are plotted in Fig. 13(a) and (b). The inductors are also modeled using the equivalent circuit of Fig. 6(b), and Table IV summarizes the model parameters based on measurement and simulation. E. Multimetal Spiral Inductors To improve the low-frequency of the spiral, we can place metal layers in shunt to reduce the series resistance of the

1478

IEEE JOURNAL OF SOLID-STATE CIRCUITS, VOL. 33, NO. 10, OCTOBER 1998

TABLE IV SIMULATED AND MEASURED MODELS OF POLYGON SPIRALS

Fig. 12.

Measured and simulated -parameters of the L4 polygon spiral.

Fig. 14. Measured and simulated -parameters of the M1M2 multimetal polygon spiral L7.

(a)

(b) Fig. 13. (a) Extracted inductance of polygon spirals from measurement and simulation. (b) Extracted -factor of polygon spirals from measurement and simulation.

spiral. To test this, the spiral L6 in the polygon set was used. Three new inductors L7L9 were constructed. L7 has two L6 spirals on metal 2 and metal 1 connected in shunt with as many vias as possible; L8 also has two L6 spirals in shunt, but with vias placed only at the ends of the segments. Last, L9 has all metals connected in shunt (except where it was necessary to use metal 0 to connect to the inner turn of the spiral). M1 spiral, and meaSimulations were run for the M2 sured and simulated -parameters are shown in Fig. 14. Good agreement is found as before. Fig. 15(a) plots the extracted series inductance of the spirals. Due to the tight coupling, the inductance value of the metalmetal structure does not drop of the spirals. The much. Fig. 15(b) plots the extracted enhancement of the metalmetal structure low-frequency is due to the drop in series resistance, in accordance with connecting the appropriate metal layers in shunt. At high frequency, though, the picture changes greatly due to substrate of the single metal layer is loss. At high frequency, the actually better. Fig. 16 compares the measured of L6L9. The measurement results indicate an overall -increase at low frequency for the M1 M2 cases, but using three metal layers is actually harmful at all frequencies (since the M0 layer is very close to the substrate). Also, the of L8 is slightly higher than the of L7, demonstrating that use of ample vias helps to boost the in the shunt conguration.

NIKNEJAD AND MEYER: SPIRAL INDUCTORS AND TRANSFORMERS

1479

(a)

Fig. 17. Measured and simulated -parameters of the coupled spirals.

Fig. 18. Circuit for calculating the power delivered to a resistive load from a reciprocal two-port network.

(b) Fig. 15. (a) Extracted inductance of polygon spirals L6 and L7. (b) Extracted quality factor of polygon spirals L6 and L7.

substrate coupling can be a signicant problem. For instance, in any amplication stage, the substrate coupling can act as parasitic feedback, lowering the gain and possibly causing oscillations to occur. Hence, it is very important to model the substrate coupling. The -parameters of two eight-turn square spirals separated by a distance of 100 m were simulated and measured. Fig. 17 shows the magnitude of the measured and simulated parameters. As can be seen from the gure, simulation results predict the coupling behavior accurately, such as the minimum To gain further insights into the coupling, we plot the power isolation from one spiral to the other using the following equations. For the arbitrary passive two-port shown in Fig. 18, through the ratio of the power delivered to a resistive load the two-port can be shown to be (22) is the current gain through the In the above equation, is the impedance looking into the two-port two-port and from the source side (23)

Fig. 16. spirals.

Measured

Q-factor

enhancement of multimetal shunt connected

(24) Using the above equations, we plot the measured and simulated power isolation for the coupled inductors in Fig. 19, where a 50- load resistance is used in the above equations. Clearly, there are two frequencies where the isolation is

F. Coupled Spiral Inductors Many RF IC designs incorporate several spiral inductors on the same die. Since these structures are physically large,

1480

IEEE JOURNAL OF SOLID-STATE CIRCUITS, VOL. 33, NO. 10, OCTOBER 1998

Fig. 21. Measured and simulated s-parameters of planar transformer. Fig. 19. Measured and simulated isolation between the coupled spirals.

Fig. 20.

Layout of planar transformer. Fig. 22. Equivalent circuit model for planar transformer.

maximum. These frequencies depend on the geometrical layout of the spirals, and this gives the designer the powerful option of placing the spirals in locations to maximize isolation. G. Planar Transformers To test the simulation accuracy for transformers, several planar transformers were fabricated and measured. An example transformer structure is shown in Fig. 20. The transformer is made of two interwound spirals each of ve turns of 7- mwide metal with a spacing of 17 m. Measured and simulated -parameters are shown in Fig. 21. Again, good agreement is found between simulation and measurement. We can gain further insight into the measurement results by using (22) to plot the measured loss of the transformer as a function of load

resistance. One can nd the optimal value of load resistance at the frequency of interest to minimize losses. The dynamics of the transformer can be captured in a compact model similar to the coupled inductors, as shown in Fig. 22. The factor for planar inductors is about 0.70.8. . While the Substrate coupling is again modeled with circuit of Fig. 22 is physically based, one can also derive the equivalent circuit, which contains an ideal transformer at the core with parasitic elements [7], [19]. The model parameters from measurement and simulation appear in Table V. V. CONCLUSION In this paper, we presented techniques to analyze, model, and optimize spiral inductors and transformers on the Si

NIKNEJAD AND MEYER: SPIRAL INDUCTORS AND TRANSFORMERS

1481

TABLE V SIMULATED AND MEASURED MODELS OF PLANAR TRANSFORMER

substrate. The techniques are accurate, taking into account substrate coupling, current constriction, and proximity effects. The analysis is also fast and efcient, making it suitable for computer optimization. Furthermore, the analysis is general and appropriate for analyzing any arbitrary arrangement of conductors, such as multimetal spirals. A custom CAD tool called ASITIC was developed that incorporates the algorithms discussed in this paper. ASITIC was used to analyze a wide variety of test structures such as square spirals, polygon spirals, coupled spirals, and transformers. The test spirals were also fabricated, and measurement results compared well to simulation. Compact models for the various devices were presented that model the device dynamics over a wide frequency range. ACKNOWLEDGMENT The authors thank W. Mack, J. Eisenstadt, and Y. Nguyen of Philips Semiconductors for their help and support in fabricating and measuring the spirals. Also, the authors thank R. Gharpurey of Texas Instruments for his valuable insights. REFERENCES
[1] N. M. Nguyen and R. G. Meyer, Si IC-compatible inductors and LC passive lters, IEEE J. Solid-State Circuits, vol. 27, pp. 10281031, Aug. 1990. [2] K. B. Ashby, W. C. Finley, J. J. Bastek, S. Moinian, and I. A. Koullias, High inductors for wireless applications in a complementary silicon bipolar process, in Proc. Bipolar and BiCMOS Circuits and Technology Meeting, Minneapolis, MN, 1994, pp. 179182. [3] J. N. Burghartz, M. Soyuer, and K. Jenkins, Microwave inductors and capacitors in standard multilevel interconnect silicon technology, IEEE Trans. Microwave Theory Tech., vol. 44, pp. 100103, Jan. 1996. [4] R. B. Merrill, T. W. Lee, H. You, R. Rasmussen, and L. A. Moberly, Optimization of high integrated inductors for multi-level metal CMOS, IEDM, 1995, pp. 38.7.138.7.3. [5] L. Zu, Y. Lu, R. C. Frye, M. Y. Law, S. Chen, D. Kossiva, J. Lin, and K. L. Tai, High -factor inductors integrated on MCM Si substrates, IEEE Trans. Comp., Packag., Manufact. Technol. B, vol. 19, pp. 635643, Aug. 1996. [6] J. Y.-C. Chang and A. A. Abidi, Large suspended inductors on silicon and their use in a 2-m CMOS RF amplier, IEEE Electron Device Lett., vol. 14, pp. 246248, 1993. [7] J. R. Long and M. A. Copeland, The modeling, characterization, and design of monolithic inductors for silicon RF ICs, IEEE J. Solid-State Circuits, vol. 32, pp. 357369, Mar. 1997. [8] J. Craninckx and M. Steyaert, A 1.8-GHz low-phase-noise CMOS VCO using optimized hollow spiral inductors, IEEE J. Solid-State Circuits, vol. 32, pp. 736745, May 1997. [9] D. Lovelace and N. Camilleri, Silicon MMIC inductor modeling for high volume, low cost applications, Microwave J., Aug. 1994, pp. 6071.

[10] S. Ramo, J. R. Whinnery, and T. Van Duzer, Fields and Waves In Communication Electronics, 3rd ed. New York: Wiley, 1994, pp. 324330. [11] A. Niknejad, R. Gharpurey, and R. G. Meyer, Numerically stable green function for modeling and analysis of substrate coupling in integrated circuits, IEEE Trans. Computer-Aided Design, vol. 17, pp. 305315, Apr. 1998. [12] A. E. Ruehli and H. Heeb, Circuit models for three-dimensional geometries including dielectrics, IEEE Trans. Microwave Theory Tech., vol. 40, pp. 15071516, July 1992. [13] F. W. Grover, Inductance Calculations. Princeton, NJ: Van Nostrand, 1946 (also New York: Dover, 1954. [14] W. T. Weeks, L. L. Wu, M. F. McAllister, and A. Singh, Resistive and inductive skin effect in rectangular conductors, IBM J. Res. Develop., vol. 23, pp. 652660, Nov. 1979. [15] R. Gharpurey, private communication. [16] C. P. Yue and S. S. Wong, On-chip spiral inductors with patterned ground shields for Si-based RF ICs, in Symp. VLSI Circuits Dig., June 1997, pp. 8586. [17] C. P. Yue, C. Ryu, J. Lau, T. H. Lee, and S. S. Wong, A physical model for planar spiral inductors on silicon, in Int. Electron Devices Meeting Tech. Dig., Dec. 1996, pp. 155158. [18] R. Groves, K. Stein, D. Harame, and D. Jadus, Temperature dependence of Q in spiral inductors fabricated in a silicon-germanium/BiCMOS technology, in Proc. 1996 Bipolar/BiCMOS Circuits and Technology Meeting, New York, 1996, pp. 153156. [19] D. O. Pederson and K. Mayaram, Analog Integrated Circuits for Communications. Norwell, MA: Kluwer, 1991, pp. 183184.

Ali M. Niknejad (S92) was born in Tehran, Iran, on July 29, 1972. He received the B.S.E.E. degree from the University of California, Los Angeles, in 1994 and the masters degree in electrical engineering from the University of California, Berkeley, in 1997, where he is now pursuing the Ph.D. degree. Mr. Niknejad has held several internship positions in the electronics industry. In the summer of 1994, he was with the Hughes Aircraft Advanced Circuit Technology Center, Torrance, CA, investigating wide-band current feedback op-amp topologies. During the summer of 1996, he was with Texas Instruments, Dallas, TX, where he studied substrate coupling. During the summer of 1997, he was with Lucent Technologies (Bell Labs), Murray Hill, NJ, where he investigated power amplier topologies for wireless applications. His current research interests include high-frequency electronic circuit design, modeling of passive devices and substrates coupling, digital wireless communication systems, numerical methods in electromagnetics, and radio-frequency computer-aided design.

Robert G. Meyer (S64M68SM74F81) was born in Melbourne, Australia, on July 21, 1942. He received the B.E., M.Eng.Sci., and Ph.D. degrees in electrical engineering from the University of Melbourne in 1963, 1965, and 1968, respectively. In 1968, he was an Assistant Lecturer in electrical engineering at the University of Melbourne. Since 1968, he has been with in the Department of Electrical Engineering and Computer Sciences, University of California, Berkeley, where he is now a Professor. His current research interests are highfrequency analog integrated-circuit design and device fabrication. He has been a Consultant on electronic circuit design for numerous companies in the electronics industry. He is a coauthor of the book Analysis and Design of Analog Integrated Circuits, (New York: Wiley, 1993), and is editor of the book, Integrated Circuit Operational Ampliers (New York: IEEE Press, 1978). Dr. Meyer was President of the IEEE Solid-State Circuits Council. He was an Associate Editor of the IEEE JOURNAL OF SOLID-STATE CIRCUITS and IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS.