CMOS Analog Filter Design for Very High Frequency Applications

: A design strategy for the synthesis of high-selectivity/low-order analog ﬁlters in Complementary Metal-Oxide-Semiconductor (CMOS) technology for very high frequency (VHF) applications is presented. The methodology for the reconstitution of a given transfer function by means of Signal Flow Graphs (SFG) manipulation in canonical form is proposed leading to a fully differential gm - C biquad ﬁlter. As a practical example, the design of a notch ﬁlter intended to suppress interferers in the lower sideband (400 MHz) of the Medical Implant Communication Service (MICS), in single-poly, 6-metal layers; Mixed-Signal/RF 0.18 µ m CMOS technology is realized. To compare the performance of the proposal with some other solution, the design of a 7th order elliptic notch ﬁlter based on Frequency Dependent Negative Resistors (FDNRs) was also accomplished. The attained simulation results prove that the proposal is competitive compared to the FDNR solution and some other state-of-the-art ﬁlters reported in the literature. The most salient features of the proposed notch biquad include: the selectivity, whose value is comparable to that of a 7th order elliptic approach and some other 3rd order ﬁlters; a high-frequency operation without resonators; linearity, with a +15 dBm IIP 3 ; a reduced form factor with a total occupied area of 0.004282 mm 2 and mostly a low design complexity. Author Contributions: Conceptualization, L.A.S.-G., C.M.-M., J.M.M.-P., C.S.-L., L.d.C.G.-P., A.L.-R., and A.I.B.-C.; methodology, L.A.S.-G., C.M.-M., J.M.M.-P., C.S.-L., L.d.C.G.-P., A.L.-R., and A.I.B.-C.; validation, L.A.S.-G., C.M.-M., J.M.M.-P., C.S.-L., L.d.C.G.-P., A.L.-R., and A.I.B.-C.; formal analysis, L.A.S.-G., C.M.-M., J.M.M.-P., C.S.-L., L.d.C.G.-P., A.L.-R., and A.I.B.-C.; investigation, L.A.S.-G., C.M.-M., J.M.M.-P., C.S.-L., L.d.C.G.-P., A.L.-R., and A.I.B.-C.; writing—original draft preparation, L.A.S.-G., C.M.-M., J.M.M.-P., C.S.-L., L.d.C.G.-P., A.L.-R., and A.I.B.-C.; writing—review and editing, L.A.S.-G., C.M.-M., J.M.M.-P., C.S.-L., L.d.C.G.-P., A.L.-R., and A.I.B.-C.


Introduction
The most remarkable development in filter theory dates from the early decades of the last century [1][2][3]. As signal processing in electronics engineering became more relevant, this branch of knowledge evolved rendering filter structures capable of accomplishing functions whose theoretical background is rigorous and elegant at the same time. Examples of such filter architectures are digital filters. Digital filtering is employed in a vast cluster of applications such as voice encoding/decoding [4], image processing [5], control systems [6], data compression [7], and telecommunications [8], to name a few. While many filtering tasks use digital signal processing, continuous-time filters are still important. Some of the assignments that analog filters can carry on include: frequency duplexing in radar and

Q Enhanced gm-C Filters
The concept of Q is used in many different contexts, and its fundamental definition relates total stored energy to energy loss [20]. On the scene of filter theory, it refers to the ability to pass a set of signals within a bandwidth while rejecting the waveforms outside the band pass. A common manner for increasing the Q of an integrated filter without augmenting the filter order is by means of tank circuits. Such approach presents the advantage of allowing the implementation of filters at very high frequencies [15]. However, the main drawback of active filters based on tank resonator circuits is the fact that the range at which the cutoff frequency can be tuned is rather narrow. Moreover, the implementation of high quality passive devices such as inductors in a Silicon technology is rather complicated [20].
On the other hand, the proposed procedure herein consists of manipulating the SFGs of a biquad gm-C filter in order to modify, merely, the first order coefficient from the denominator polynomial of its transfer function and hence to improve the selectivity of the circuit.

Formulation of the SFG from a Filter Transfer Function
Typically, when synthesizing active filters by means of SFGs, the starting point is to build up the graph of a passive ladder network, commonly a filter, and then to proceed with an active leapfrog realization [21]. Another way of using the SFGs approximation is to draw the graph directly from the transfer function of the filter [22].
Let the transfer function of a filter be (The reason because the form (1/s) is employed instead of s in (1) is that integrators are preferred over differentiators in a practical circuit realization.) [22] T(s) = a m 1 s n−m + a m−1 1 s n−m+1 + · · · + a 1 1 s n−1 + a 0 1 s n 1 + b n−1 1 s + · · · + b 1 1 s n−1 where m, n ∈ Z satisfying m ≤ n, and a i , b j ∈ 1 , are the i-th and j-th coefficients of the numerator and denominator polynomials, respectively. It can be proved by means of the gain formula of Mason that the SFG in controllable canonical form of Figure 1a describes the transfer function expressed in (1). The reason of using such graph is that it has two important features: all loops in the graph touch each other and every forward pathway from the input node to the output node touches all loops. These characteristics allow for enunciating the gain formula of Mason as [22] where P k is the k-th forward path weight, L n is the n-th loop weight, and summations are all over the direct pathways as well as the loops. Thus, L 1 = − b n−1/s, L 2 = − b n−2/s 2 , · · · , L n = − b 0/s n , whereas P 1 = a m , P 2 = a m−1/s, · · · , P n = a 0/s n . The convenience of the graph representation in the form depicted in Figure 1a lies in its feasibility for adjusting a specific coefficient, either (or even both) from the numerator or denominator polynomials of T(s) by changing the value of the corresponding forward/backward pathway in a given loop. This is demonstrated in the successive.

Synthesis of gm-C Biquad Filters Based on SFGs
If both numerator and denominator polynomials of (1) fulfill the condition that m = n = 2, then T(s) corresponds to the biquad filter case, i.e., Hence, the low-pass (a 2 = a 1 = 0), high-pass (a 1 = a 0 = 0), band-pass (a 2 = a 0 = 0), band-stop (a 1 = 0) or all-pass responses can be performed. Figure 1b illustrates the SFG of (3) when the input and output nodes are in voltage mode.
In order to compose the SFG of the biquad structure in (3) from a gm-C filter approach, the following changes must be made to the graph in Figure 1b: • the forward pathway in every single loop has to be pulled apart into two twigs, one with a path weight gm j , and the other with a path weight (1/sC j ). The subscript j denotes the j-th transconductance and capacitive reactance along the straight path from node input to node output. Namely, the integrator on the graph is compounded by the product of a transconductance and the reactance of a capacitor. • the split-up of the j-th forward pathway within the loops along the graph is achieved by the inclusion of a current node, I k with k = 1, 2, 3, . . . , between the voltage nodes which surround the signal flow on the forth path. Figure 2a sketches the resulting SFG for the gm-C biquad filter structure obtained by means of the procedure formulated above. When applying the gain formula of Mason in Equation (2) to the graph, it is obtained For simplicity, we take a 2 = a 1 = a 0 = b 0 = 1, gm 1 = gm 2 = gm, and C 1 = C 2 = C. Such simplification is illustrated in Figure 2b. Thus, Herein, the coefficients of the zero and the first order from denominator polynomial are defined as [14] gm where f 0 and Q are the center frequency and the quality factor of the filter, respectively. The implementation of the SFG of Figure 2b may be accomplished easily by using the building blocks, in fully differential mode, depicted in Figure 3, where the components of the graph, i.e., the integrators, the feedback between two nodes, the signal summation at a node, and the direct pathway between two voltage nodes are posed.

Increasing Selectivity by SFG Manipulation
By scrutinizing the ratios given in (6), it can be appreciated that selectivity is inversely proportional to b 1 . In this way, as this value decreases, Q rises monotonically and vice versa. Here is where the SFG representation pays off. By inspection of the graph in Figure 2b, it can be seen that the loop gain formed by gm, 1/sC, and −b 1 , among the nodes V in , I 1 , and V 1 , establishes the first order coefficient from the denominator polynomial of the transfer function expressed in (5). In order to modify the Q of the filter without changing the center frequency, f 0 , neither the value of the transconductance gm nor the capacitance C must be altered. Consequently, −b 1 is the only parameter that is allowed to adjust. Manipulation of the loop can be done to redefine Q. Figure 4a exposes the suggested SFG by which the Q of the biquad gm-C filter can be changed independently of f 0 . It can be seen that a new loop between the nodes I 1 and V 1 (dashed lines in Figure 4) has been composed, substituting that formed by branches gm, 1/sC, and −b 1 . The gain of the new loop is given by branches 1/sC and −gm f . By using (2), the transfer function of the new graph is found: and now f 0 and Q are expressed as (c) current addition at a node and its conversion to a voltage node; (d) a direct pathway with unity gain between two voltage nodes. Therefore, f 0 is preserved as in the former case, whereas Q is now the ratio of transconductances gm and gm f . Thus, Q can be tuned within a range by fixing gm to a given value, which is determined by the center frequency of the filter, and varying gm f . From (8), it is clear to see that to achieve high Q values, gm f must be smaller than gm. At this point, it is important to inquire how large the ratio (gm/gm f ) can be. Numerically, it is possible to choose a ratio as large as 100, for example, but in practice there are limits which make unfeasible that possibility. Since gm is fixed according to f 0 , gm f is the only term in (8) with degree of freedom to be varied. However, the transconductance of any analog architecture in a CMOS process is somehow related to the aspect ratio, (W/L), of the transistors that compose it. In fact, the transconductance rises or decreases as (W/L) does in either a linear or quadratic fashion [23]. Therefore, a small value of gm f implicates a small (W/L) ratio of the transistors involved. There are limits concerning the minimum (W/L) ratio in a CMOS design imposed by the design rules of a given process. Consequently, gm f can be smaller than gm to some extent only. Furthermore, if the current of gm f at its output port is delivered to a capacitor load, which is the case for the gm-C filter, then how fast capacitor is charged and discharged plays an important role. The time constant, τ, in gm-C integrators is given by the (C/gm) ratio. For the case of f 0 = 400 MHz, the time period of the signal is T 0 = 2.5nsec. In case that τ is established, in the worst case, at 40% of T 0 , then τ = 1nsec. Considering that C is in the hundred of femto Farads, then gm f must be at most in the dozens of µ . Hence, the lower limit for the value of gm f depends on f 0 as well.
Altogether, a considerable Q enhancement with the proposed biquad filter synthesis based on SFG manipulation can be achieved compared to the selectivity exhibited by a regular biquad approach. Simulation results at transistor level of a gm-C structure synthesized by means of the proposal demonstrate that biquad architectures with performance metrics even better than higher order filter structures are conceivable. A band-stop response has been chosen with the aim of realizing the synthesis of a biquad RF notch filter for blocker ( Figure 4b). Nevertheless, the Q exhibit the same behavior for any of the approximations (low-pass, high-pass, band-pass, all-pass) since the denominator polynomial of the transfer function of the biquad filter remains the same for all the different responses.

An RF Biquad Notch Filter for Jammer
Due to the huge number of wireless devices sharing a given portion of the electromagnetic spectrum and operating in close proximity, a healthy coexistence among diverse communication standards has become a relevant issue. It is possible to have interferers or blockers (also called jammers) as strong as 0dBm, driving almost any receiver in compression [24]. Therefore, RF jammer filtering is needed.

The biquad gm-C Notch Filter Derived from SFG Synthesis
Since fixed filters are undesired because of their limited Q and tunability, the Q-enhanced bandstop gm-C biquad filter in Figure 5 is proposed, as a design example, for blocker suppression at f 0 = 400 MHz, which is the lower-sideband of the Medical Implant Communication Service (MICS), whose frequency band lies within the (402-405) MHz [25]. Such filter architecture is built up directly from the SFG of Figure 4b for the notch case by interconnecting the fully differential building blocks of Figure 3. By doing this, the following transfer function is produced: whose stop frequency, f 0 , and selectivity, Q, are expressed as in (8). k is the scaling factor from the kC capacitors in Figure 5. It is preferable that k >> 1, otherwise, f 0 deviates from its nominal value. This is depicted in Figure 6. As can be appreciated, for k = 1, f 0 experiences a deviation of ∼ 42% of its nominal value; furthermore, when k = 10, f 0 deviation is ∼ 5% of its nominal value; finally, when k = 100, f 0 deviation is negligible and it remains at ∼ 400 MHz. Thus, by taking k = 100, the transfer function of the notch filter of Figure 5 becomes To achieve operation at the hundreds of MHz, the transconductance gm has to be on the few m while the capacitor values, C, must prevail on the hundreds of femto Farads (parasitics included). As stated earlier, to exhibit a high Q value, transconductance gm f of the Filter has to be, at most, on the dozens of µ . These values can be easily managed in the single-poly, 6-metal layers, Mixed-Signal/RF 0.18 µm CMOS technology that was employed for the design of the gm-C biquad.  A simple transconductor topology with a minimum number of internal nodes and consequently with a high frequency of operation was proposed by Bram Nauta in [16]. Even though its major drawback of drawing transient currents from the power supply, which increases the potential of signal coupling between stages [26] as well as exhibiting a considerable power consumption at very high frequencies, with more than 300 citations, papers and patents included, the circuit has proved to be a very useful building block. Figure 7 shows the Nauta transconductor with CMOS devices. As can be seen, it consists of a fully-differential amplifier based on six inverter gates, i.e., six PMOS (M PG1,...,PG6 ) and six NMOS (M NG1,...,NG6 ) transistors. In order to size the transistors of the transconductor, the design trade-offs among important parameters such as low frequency gain, distortion, signal swing available, dissipation and frequency limitations must be considered according to the recommendations in [14]. By doing so, and aiming at transconductance values of (gm/2) = 500 µ and (gm f /2) = 25 µ , the following aspect ratios arise: ( W M NG1,...,NG6 /L)=( 14.04 µm /0.18 µm) and ( W M PG1,...,PG6 /L)=( 42.3 µm /0.18 µm). The C capacitors are meant to be on chip with a capacitance of 200 fF (parasitics included); meanwhile, the kC capacitors are chosen to be off-chip with a capacitance of 20 pF. The bias voltage, V dd , is set to a value 1.8 V. The layout of the filter is shown in Figure 8; it was done in the CADENCE IC6.1.6.101 Virtuoso design environment. The total occupied area of the filter is 0.004282 mm 2 , with a length of 86.61 µm and a width of 49.44 µm. Capacitors, C, are Metal-Insulator-Metal (MIM) Capacitors. In Section 4, the post layout simulation results of the proposal are reported.

The Alternative Filter Solution Based on Active Simulation of Passive LC Networks
Before presenting the performance features of the proposed CMOS band-stop biquad filter, it is important to point out that there are systematic approaches other than the structural simulation method based on the use of the SFGs to synthesize active filters [27]. One possibility is to break the desired transfer function, obtained by some known approximation method like Butterworth, Chebyshev, Cauer, Bessel-Thomson, or Pascal, into the product of first and second order transfer functions with either real or complex conjugate poles; then, these low-order transfer functions are implemented with active structures through the use of operational amplifiers (OPAMPs), transconductance amplifiers (OTAs), Current Conveyors (CCs), etc.; finally, the active structures are cascaded to obtain the wanted transfer function. Probably, the most popular synthesis methodology of active filters is the one based on active simulation of passive LC networks with double resistive termination. Such approach can be done by performing the direct replacement of inductors in the LC ladder network by general impedance converters (GICs) or the scaling of frequency dependent impedances in the ladder network. In order to compare the suggested SFG based notch biquad filter with other solutions, the design of a fllter based on active simulation of passive LC networks with double resistive termination was carried out.   Figure 9. Cont.

Post Layout Simulation
Post layout simulation of the proposed biquad notch filter was performed with the aid of the Cadence IC6.1.6.101 Spectre analog design environment. Furthermore, simulation of the FDNR active filter of Figure 9b in active form was also carried out in the single-poly, 6-metal layers, Mixed-Signal/RF 0.18 µm CMOS technology. The attained results are summarized in Table 1 along with some other state-of-the-art band-stop filters. From those, the proposals [28][29][30] report simulation results based on CMOS circuits, whereas approach [31] accounts for experimental results based on lumped resonators. The latter is the filter whose center frequency, f 0 , is the larger, closely followed by the biquad and FDNR filters presented in this work. The f 0 of the rest of the approaches are at least one order of magnitude smaller. In addition, since acoustic-wave-lumped resonators are employed in [31], this is the proposal with the larger Q of all the filters in Table 1, at the expense of a null tunability. The applications for which the filters reported somewhere else are referred only in [29] (Bluetooth), the rest do not specify where the design efforts were directed to. On the other hand, both the SFG based and the FNDR filters reported herein are intended for the Medical Implant Communication Service (MICS), whose frequency band lies within the (402-405) MHz [25]. Figure 10 shows the frequency response of the magnitude of both the suggested SGF CMOS biquad and the 7th order FDNR. As can be seen, these achieve a maximum attenuation of ∼−17 dB at the notch frequency (≈400 MHz); however, the FDNR filter exhibits a narrower BW (≈13 MHz) compared to the SFG biquad (BW ≈ 20 MHz) but with a larger power consumption and filter complexity. This was expected since a higher filter order with ripple in the bandpass renders a higher selectivity; however, the reached selectivity (Q = 20) of the proposed architecture is a good quality factor for a biquad, even though the NAM based biquad in [28] reports a Q of 50 for a rather low frequency of operation f 0 = 2 KHz. A Q = 20 was also achieved in [29]; nevertheless, a third order approach was used and hence the circuit possesses a steeper slope (20 dB/Dec sharper). Otherwise, the notch depth reported in [28][29][30] is optimistic; even with a very narrowband like those in [31], the achieved rejection barely surpasses the −20 dB. Thus, the −17 dB accomplished by the two circuit topologies described here is more realistic; it is not uncommon to have RF jammer filters with a notch depth barely larger than −10 dB.   In terms of tunability, the filter in [29] is comparable to the SFG biquad and the FDNR; however, these work at a higher frequency. Unfortunately, none of the filters in Table 1 reported somewhere else  present distortion measures (THD, I M 3 , I IP 3 ); on the contrary, the SFG biquad and the FDNR reported herein exhibits a good linearity with +15 dBm and +12 dBm IIP3, respectively. Again, sensitivity is not reported in the rest of the filters, but, for our case, the SFG biquad is more sensitive compared to the FDNR, as expected since the latter active architecture is synthesized from a passive network. Finally, power consumption is the drawback of the proposal. Both the SFG biquad and the FDNR are the most power hungry among the filters in Table 1. Though there is the chance to scale the value of gm and C to some extent to lower power consumption, by doing so, maintaining the performance of the filter is troublesome. Since the dissipation of any gm-C filter is strongly determined by the dissipation of its transconductances, and for the case of the Nauta transconductor employed herein the power dissipation can be reduced by lowering the bias supply, then less V dd is favorable for the dissipation of the proposed biquad notch; unfortunately, when V dd goes down the cutoff frequency of the transconductor does as well. Actually, the results reported in [16] for the case of a third order elliptic low-pass derived from a passive ladder show that for an ≈75% decrease of V dd the power consumption is lessened ≈ 90% at the expense of an ≈77% cutoff frequency reduction. Regarding the used CMOS technology and its impact on the power consumption of the proposal, the following question arises: is it possible to significantly reduce the power dissipation of the proposed filter structure if a more advanced technology, say 65 nm, is used? The answer is that the power consumption would be moderately reduced. The reason is that gm-C filters are, typically, noise and open-loop distortion dominated circuits whose performance parameters such as the power dissipation do not show dependence on technology scaling as long as the signal-swing scales down as the supply voltage diminishes [32]. By glancing at some published works about gm-C filters and Nauta transconductors in 65 nm [33][34][35], power dissipation in the tens of milliwatts are reported, which are slightly smaller than the ≈52 mW exhibited by the proposal in 180 nm. Thus, if power consumption is an issue, it is recommended to use a gm-C topology other than the biquad employed here; for instance, a more energy efficient gm-C structure for low-order filtering is reported in [36].
It is important to remark that practical loading to the filter is beneficial in order to verify its driven capability. However, in the proposed filter structure, this was ignored since the scope of the research is to provide a methodology for synthesizing high-selectivity/low-order biquad gm-C filters by means of SFG manipulation instead of focusing on the I/O buffering needs of a practical realization. However, this can be covered with the inclusion of a driver stage, a programmable gain amplifier (PGA) for instance. Some filter designs include both the filter structure and the PGA, especially if energy efficiency is relevant [36].

Monte Carlo Simulation
On the other hand, to verify the robustness of the proposal under the presence of mismatch, Monte Carlo simulations based on the model of Pelgrom were conducted. Figure 11 shows the Monte Carlo simulation of the SFG biquad. On the left side, there is the frequency response of the filter. On the right side, at the top, there is the f 0 histogram over 100 runs. Again, on the right side, but at the bottom, there is the Q histogram over 100 runs. As can be seen, the mean is 400.7 MHz for the f 0 with a standard deviation of 1.07 MHz. Thus, the center frequency has a nominal value of ∼400 MHz with a tolerance of ±0.25%. On the other hand, the mean is 20 for the Q with a standard deviation of 5.7. Hence, deviation is larger for selectivity compared to f 0 . The mean value of Q is 20 and ∼95% of the variations of Q lie between 8. 6

Conclusions
The synthesis of high performance biquad gm-C filters by means of SFG manipulation has been presented. The SFG representation of a transfer function in canonical form is formulated along with the methodology to build up this transfer function in a fully differential gm-C circuit. As a practical example, the design of a gm-C notch biquad filter intended to suppress interferers in the lower sideband (400 MHz) of the Medical Implant Communication Service (MICS), in a single-poly, 6-metal layers, Mixed-Signal/RF 0.18 µm CMOS technology was realized. Additionally, to compare the performance of the proposed notch biquad with an FDNR solution, the design of an active 7th order elliptic notch filter was also accomplished. The attained simulation results prove that the proposal is competitive compared to both the FDNR solution and some other state-of-the-art filters reported in the literature. The most salient features of the proposed notch biquad filter include: the selectivity, whose value is comparable to the quality factor of a 7th order elliptic approach and some other 3rd order filter architectures; a high-frequency operation avoiding the use of resonators; linearity, with a +15 dBm I IP 3 ; a reduced form factor with a total occupied area of 0.004282 mm 2 ; and a low design complexity. Monte Carlo simulation based on the model of Pelgrom over 100 runs show that the center frequency of the proposal has a nominal value of ∼ f 0 = 400 MHz with a tolerance of ±0.25%; meanwhile, the mean value of the Q = 20 with ∼95% of the variations between 8.1 and 31.4. Finally, power consumption of the proposal may be troublesome in the case that low power consumption is wanted; in that case, the recommendation is to use a different gm block other than the transconductor of Nauta, whose main drawback is its considerable power consumption.