WorksEngineering
A Hybrid Digital/Analog Polyphase Filtering Approach for High-Speed RF-DACs
Abstract
This paper proposes a hybrid polyphase interpolation for RF-DACs that removes the need for full-rate digital commutation by using passive combining on the RF-DAC itself. A P-branch polyphase FIR filter is factored into M streams, each formed by a (P/M)-way digital commutator at fRF/M and driving a dedicated RF-DAC slice, giving either partial digital (1<M<P) or fully analog (M=P) recombination. We show that the slice array acts as a single full-rate DAC with an M-times longer zero-order hold, derive yield laws for slice gain, timing and unit-capacitor mismatch, and find that with retiming the binding requirement is sub-picosecond matching of the fRF clock arrival, so the best choice is the smallest M that closes digital timing. Two 28-nm digital macros, designed for a 12.8GSps output rate and verified by post-layout simulation, illustrate a 4-phase variant that closes timing up to 14.4GSps with its digital logic at fRF/4 and an 8-phase variant that closes timing up to 16GSps with its digital logic at fRF/8. The analog recombination is evaluated with a behavioral model of the slice array. The architectures target wideband software defined radio (SDR) platforms, including cognitive radio, electronic warfare, and satellite communications.
Introduction
Wideband software defined radios (SDRs) are central elements for modern wireless communications systems, including those in commercial and military usage. They enable frequency-agile transceivers that can accommodate a wide variety of waveforms from single hardware platforms [1]. SDRs have become relatively commonplace below 7 GHz (frequency range 1 (FR1)) [2], owing to the increasing sampling rate that is available for data converters. The advent of the radio-frequency digital-to-analog converter (RF-DAC) is a primary enabler for wireless transmitters (TXs), in that it can directly synthesize baseband (BB) waveforms and output them directly at radio frequency (RF) [3, 4, 5].
However, in both commercial and military networks, higher frequency spectrum, such as that used in both frequency range 2 (FR2) and frequency range 3 (FR3) bands is being proposed to offset the severe congestion that exists in FR1 bands [6]. A key challenge to scaling hardware systems to higher frequency spectrum is that the BB signals typically use
All authors are with the Bradley Department of Electrical and Computer Engineering, Virginia Tech, Blacksburg, VA 24061 USA. (Corresponding author: Jeffrey S. Walling, e-mail: [email protected]). This work was supported by the National Science Foundation (NSF) under grant #2314813. Simulation code and models available at https://github.com/noyades/Non-uniform-Sampling-RF-DAC.
wider bandwidths and in the case of RF-DACs the BB signals must be interpolated to higher sample rates to help suppress spectral images that arise from the process of sampling.
In conventional RF wireless systems (e.g., digital-BB + digital-to-analog converter (DAC) + RF front-end), the sampling images can be suppressed using active, high-order filters that serve as the reconstruction filter of the DAC. These are impractical when using an RF-DAC, where the reconstruction filter is typically a low-order passive RF filter acting as an impedance matching network. Instead, the sample rate must be increased as much as possible, and ideally as high as the output frequency, which enables easier suppression of the images. This is similar to the concept of oversampling in BB DACs.
Advanced CMOS process nodes are capable of supporting high sampling rates [7, 8, 9]. However, these are still traditional baseband architectures that require separate up-conversion chains and hence do not offer the flexibility of the RF-DAC. Furthermore, the designs feature highly customized digital architectures to meet stringent timing requirements. Time-interleaved RF-DAC architectures [10] and direct-RF TXs with embedded FIR-like impulse responses [11, 12, 13, 14, 15, 16] have explored related techniques, but remain constrained in filter reconfigurability and achieve limited tap depth.
The challenge in purely analog interpolation methods is that they provide only a limited number of filter taps. In a time-interleaved (TI) converter array [17], the number of taps equals the number of interleaved phases. In a time-approximation filter (TAF), the limitation is on the available shaping of the clock pulse [15, 16]. Additionally, the shaping on the clock pulse can have other adverse effects such as spur generation. In analog FIR filters, the limitation is on the number of individual analog processing cores [13, 14]. This limits the out-of-band (OOB) rejection and adaptability to new waveforms. Hybrid approaches, as shown in Figure 1 that merge analog and digital filtering capabilities, are an alternative approach. A single-phase RF-DAC with a digital interpolation filter would rely on the filter operating at output speeds up to the RF carrier frequency for maximal image suppression. This is practical if operation is limited to FR1 frequencies (below 7 GHz), but becomes impractical as frequencies scale into FR3 (7–24 GHz) and FR2 (above 24 GHz), where even full-custom digital FIR filters have only approached 10 GHz [18].

Figure 1. RF-DAC interpolation architectures. (a) Conventional direct-digital FIR: requires logic running at . (b) Polyphase hybrid analog-digital FIR: lanes, each driving an RF-DAC slice, reduce the maximum digital rate to ; this paper implements and .
This paper shows that the recombination of a polyphase interpolator can be split between digital commutators and the RF-DAC itself, giving either partial digital recombination followed by analog recombination, or fully analog recombination, and implements one variant of each (Figure 2, middle and right). A Farrow fractional-rate filter in software/FPGA decouples the fixed baseband data rate from the RF clock, and the multi-GHz integer-rate interpolation is decomposed into polyphase form on a dedicated mixed-signal ASIC macro. The 4-phase variant needs only four clock phases, easily derived from the RF local oscillator, but its digital logic must close timing at ; the 8-phase variant needs a more complex multiphase clock network but halves the digital rate to .

Figure 2. Detailed polyphase interpolation filter block diagrams. (left) Conventional polyphase FIR integer rate resampling, with recombination in the digital domain. Output delay and sum operations run at . (middle) 4-phase polyphase decomposition: eight sub-filters running at , before parallel branches are recombined at , each driving a dedicated RF-DAC slice. (right) 8-phase polyphase decomposition: eight sub-filters running at , each driving a dedicated RF-DAC slice.
Section II describes the generalized recombination and the system partitioning. In Section III the recombination model, along with its mismatch, clock and in-band noise and distortion budgets is analyzed. Next, Section IV provides guidance on the optimal choice of the number of clock phases, . Implementation and post-layout results are presented in Sections V and VI. Finally, Section VII concludes the paper.
System Architecture and Partitioning
The block diagram of Figure 1(b), and the detailed signal-flow graphs of Figure 2, illustrate the proposed hybrid architectures for implementing a high-rate interpolation filter for an RF-DAC output. To achieve frequency agility while maintaining physical implementation viability, the architecture employs a distinct partitioning. The system relies on an initial Farrow filter structure for efficient fractional-rate resampling of the baseband signal to a rate that is an integer fraction of the desired RF output frequency. Farrow filters are well-suited for this role due to their ability to perform arbitrary-rate interpolation with a fixed set of filter coefficients [19, 20]. Given its mathematical complexity and requirements for runtime parameter adjustment, this fractional block is designated for flexible off-chip deployment (e.g., on an FPGA front-end).
The output of this off-chip Farrow filter then feeds the custom mixed-signal ASIC macro, which houses a polyphase FIR bank that generates multiple parallel hardware output streams, each operating at a manageable fraction of the target output frequency. As shown in Figure 2 (left), a pure-digital polyphase decomposition requires a digital commutator running at to reconstruct the output sequence. The proposed hybrid architectures move all or part of this recombination into the RF-DAC, where it is performed passively. This can be done using either current- or voltage-mode RF-DACs, by summing currents [21] or by controlling charge division on capacitor arrays [22], with the recombination controlled by interleaved clock phases.
Generalized Digital/Analog Polyphase Recombination
A length- FIR filter with coefficients is decomposed into polyphase sub-filters [23, 24, 25],
where each sub-filter, with for , is defined as:
Each sub-filter runs at . The recombination in (1) is what must run at . Set , and index the sub-filters
| (digital) | (this work) | (this work) | (this work) | |
| Sub-filter rate | 1.6 GHz | 1.6 GHz | 1.6 GHz | 1.6 GHz |
| Digital commutators ways | none | |||
| Highest digital rate | 12.8 GHz | 6.4 GHz | 3.2 GHz | 1.6 GHz |
| RF-DAC slices | 1 | 2 | 4 | 8 |
| Clock phases (spacing) | — | 2 (180°) | 4 (90°) | 8 (45°) |
| Analog hold | 78.1 ps | 156 ps | 312.5 ps | 625 ps |
Table I. Rates and resources of the generalized recombination, ,
as with and . The same filter then factors into two nested recombinations,
where , the -th -phase component of , is assembled from of the sub-filters:
The sum in (4) interleaves sub-filters into one stream at : a -way digital commutator. The sum in (3) interleaves the streams produced by the into the full-rate output, and it is this stage that the RF-DAC performs, with one slice per stream and clock phases spaced apart at [22, 26]. The pair thus spans a design space whose regimes are shown in the three columns of Figure 2:
(): fully digital recombination. A single commutator runs at and drives one DAC (Figure 2, left).
: partial digital recombination. commutators run at , and their outputs are recombined in the analog domain by RF-DAC slices (Figure 2, middle, , ).
(): fully analog recombination. Every sub-filter drives its own slice with no digital commutators (Figure 2, right, ).
Table I lists the resulting rates. It is noteworthy that the digital commutators are synchronous logic whose correct timing is guaranteed by static timing analysis at sign-off, so they contribute no amplitude or timing error of their own, and every analog non-ideality (e.g., the staggered hold, the slice mismatch and the clocks of Section III) depends on alone. Additionally, the fastest digital clock required is for any , so it is , and not , that sets the required digital timing closure; only sets how parallel the FIR arithmetic is [27]. The optimal choice of is analyzed and quantified in Section IV.
Polyphase FIR Architecture
To demonstrate the effectiveness of the proposed techniques, we designed an interpolation filter targeting 802.11be-like 320 MHz signals at FR3 frequencies. Though no monolithic standard currently exists, this particular combination of wide signal bandwidth (320 MHz) and ultra-high output frequency (12.8 GHz nominal, up to 16 GSps signed off) serves to demonstrate the capabilities of the proposed technique for high-dynamic-range, wideband signals.
Filter Specification
The integer-rate interpolation from to is split into a pre-interpolator, from to , followed by a stage. The pre-interpolator is designed to a 160 MHz passband, a 240 MHz stopband, and dB stopband attenuation (before coefficient quantization). The stage keeps the same 160 MHz passband and rejects the images of the stream, centered at multiples of 1.6 GHz.
To achieve these characteristics, the stage cascades a FIR filter, which compensates the passband droop, with the response of a 6-stage cascaded integrator-comb (CIC) filter [28]. The sinc response of the CIC places its nulls at the image centers; its droop is removed by the compensation filter, which also enforces the stopband attenuation [29]. No CIC structure is built: the cascade is implemented as an equivalent direct-form FIR for polyphase decomposition.
Polyphase Decomposition and Coefficient Quantization
The coefficients of the equivalent cascade were derived using the impulse response of the full structure, quantized to 16 bits for hardware execution, and decomposed according to (1) and (2) with , giving 165 taps. The pre-interpolator is a 95-tap Nyquist(4) filter, also with 16-bit coefficients, implemented as four polyphase branches. After quantization, the pre-interpolator attenuates its stopband by 66.8 dB, and the cascade attenuates the images of the stream by more than 90 dB, both well above the 55 dBc spurious-free dynamic range (SFDR) target of Section III.
To enable a rigorous comparative analysis between the 4- and 8-phase physical macro variants, we utilize a single set of master filter coefficients and apply the same polyphase decomposition framework to both. The 4-phase architecture combines interleaved pairs of the 8-phase branches in the digital domain, i.e., (3) and (4) with , while the 8-phase architecture directly implements all eight physical hardware branches. This approach ensures both variants maintain an identical underlying filter response, isolating the post-layout evaluation to differences in digital logic power, clock routing, and layout congestion.
For the hardware instantiation, the input bitstream data is fed in parallel to all eight sub-filters on the macro, each operating at 1.6 GHz and outputting one sample per clock cycle. These eight outputs map to successive time offsets (, where ). These parallel streams directly drive the input ports of the eight RF-DAC slices, sequencing their updates across the 45°-spaced clock phases to passively reconstruct the 12.8 GHz analog output waveform.
Architecture Trade-offs
A single FIR at GHz (Figure 1(a)) exceeds what even full-custom digital FIR filters have achieved; these have only approached 10 GHz, by relying on power-of-two coefficients, bit-level pipelining and supply overdrive [18]. A fully digital polyphase decomposition () still needs an output commutator at . Both implemented variants avoid these constraints. The 4-phase variant (, ) needs four 90°-spaced phases at , which a divide-by-4 counter clocked by provides directly, but its four two-way commutators must close timing at 3.2 GHz. The 8-phase variant () needs eight -spaced phases, a more complex clock network, in exchange for removing the commutators, so its fastest digital clock is the 1.6 GHz sub-filter rate.
Recombination Model and Mismatch Effects
Staggered Zero-Order Hold
This section analyzes the analog stage of (3), the recombination of the streams on the RF-DAC. As argued in Section II-A, the digital commutators contribute no error of their own, so every result below depends on alone and holds for any , i.e., for partial and fully analog recombination alike.
The recombination performed on the slice array is not a one-hot commutator. On each slice update-clock edge all stream words are held, then loaded into the array sequentially, one per successive cycle. Once loaded, a slice continues to switch at with that data until its next update, so every slice drives the array for a full slice update period . If the slice weights are equal, the slices differ only in when their data changes, and the array output is a staggered zero-order hold,
where is the baseband-equivalent output, is a -wide unit pulse, and is the stream that drives slice , i.e., the output of in (4) with . Since , where is the output of the full-rate filter , substituting in (5) gives
Hold Response
By (6), the array is equivalent to a single DAC updated at whose zero-order hold lasts instead of [30, 31]. Its frequency response is the usual sinc of a zero-order hold, , but with nulls at multiples of rather than of ; at RF they lie at from the carrier. The -times narrower sinc also increases the passband droop, but only to 0.04 dB () and 0.14 dB () at the 160 MHz band edge, which the FIR can pre-compensate if required. A sample-level simulation of the slice array agrees with (6) to within , i.e., double-precision rounding, for both variants.
Effective Resolution
(6) also bounds the achievable resolution. Slices carrying successive time samples add full-scale output and update rate but never subdivide the unit element, so slices of bits do not yield bits. Each slice quantizes independently, so the in-band signal-to-noise and distortion ratio (SNDR) improves over a single slice by only , i.e., bits (+1.0 bit for , +1.5 bit for ), which simulation confirms. The array therefore falls bits short of a ()-bit DAC, even with perfectly matched slices. Section III-B treats unequal slice weights and timing skew, whose spurs at fall on the skirts of the sinc nulls.
Impact of Timing and Gain Mismatch on Spurious Outputs
With a per-slice gain error and timing error/skew , the output spectrum is
where is the branch spectrum, that is periodic with , and . Because every slice holds for exactly , factors out of the sum. Hence, mismatch never moves the location of the null, it only reduces the cancellation it achieves across the null’s skirts.
For a single tone at each is a comb at . Within of the carrier, the output has only the carrier () and the polyphase spurs , . With , the hold weights the spurs relative to the carrier by the exact ratio:
To a first order in and , the spur-to-carrier amplitude is then:
where , , and and are the -point discrete Fourier transforms (DFTs) of the slice gain and timing errors [32, 33]. A skew gives slice a phase error of at the spur frequency , and the hold attenuates that spur by from (8). Since , the product is for every , so the timing sensitivity depends on the signal frequency and not on [34]. Skew of the clock itself behaves differently, because it also rotates the carrier of the slice; Section III-C shows that its per-spur weight is times larger, i.e., 12.5 dB at and 19.0 dB at , which tightens the median skew budget by 10.6 and 14.9 dB (Table II). The gain error is attenuated by alone, least at , where is smallest.
To turn (9) into a budget, let the slice errors be independent, zero-mean Gaussian random variables. The gain (or weighting) error is , while the timing error is , where and are the respective standard deviations of the errors. For a real white Gaussian sequence, the DFT bins are mutually independent, and bin is the complex conjugate of bin . The spurs at therefore have the same random amplitude, and we refer to them as group . For the bin is a circular complex Gaussian of variance , so the spur power is exponentially distributed. The bin is real, so its spur power is a scaled variable. Taking the larger spur of each group, the mean spur power of group is
and the two mechanisms add in power. Timing skew contributes equally to every group, whereas gain error falls mainly on . The SFDR is set by the worst spur. Because the groups are independent, its cumulative distribution is the product of the group distributions,
For a single mechanism, every is proportional to , where is either or . Let denote the yield, i.e., the fraction of fabricated chips whose worst spur lies below a level , so that . This level can then be written as , and a yield at an SFDR target of dBc is achieved if [35]:
The yield cost in SFDR is the gap between and the largest group mean . For gain error, dominates and the worst spur behaves like a single exponential variable: it lies 1.1–1.4 dB below at the median and 4.8 dB above it at 95% yield. For timing skew, all groups have the same mean, so the worst of them grows with : 0.7 and 2.7 dB above at the median for and 8, and 6.4 and 6.9 dB at 95%.
A given SFDR target follows from the slice resolution of the subsequent RF-DACs. In our example each slice is a sign-magnitude DAC with an 8-bit magnitude (255 unary units plus sign, i.e., 511 levels), whose quantization-limited signal-to-noise ratio (SNR) for a full-scale sine is 55.9 dB. Recombining slices adds , as shown in Section III-A2, i.e., 61.9 dB at and 64.9 dB at , over the Nyquist band for a full-scale sine. Hence, a reasonable limit for the SFDR target is set by the quantization-limited SNR of a single slice. For example, mismatch spurs at or below 55 dBc remain below the quantization noise of any one slice. For a 55 dBc SFDR with the tone at the band edge, MHz, (13) gives a median budget of ps and at , and ps and at . At 95% yield the budgets tighten to 1.70 and 2.27 ps, and 3.9% and 2.6%. For a different target , e.g., with higher-resolution slices, the error budgets scale by , i.e., they halve for every 6 dB of additional SFDR. To verify the closed-form estimates, a -trial Monte Carlo of the exact, non-linearized (7) was run, and it agrees with them to within 0.1 dB for every SFDR above 45 dBc; beyond ps the first-order expansion becomes conservative, by at most 1.25 dB at 20 ps. The yield law also explains why doubling relaxes the median timing budget by only 12% rather than the that the single-spur mean (10) suggests: the reduction of each spur is largely offset by the larger number of independent spurs competing for the worst case. The gain budget tightens with because doubles.
In the case of a modulated stimulus of bandwidth , each spur is spread over the band, at baseband offsets between and . Averaging over a flat band gives , so to leading order in (10) and (11) is replaced by , and the spur power drops by a further factor of 3. For gain this is 25–31 dB below the conventional interleaving result . A Monte Carlo simulation with a 320 MHz modulated stimulus, evaluated on the worst spur band, gives budgets of 5.7 ps () and 6.6 ps () for timing skew and 14.5% and 10.0% for gain error. These are all looser than the single-tone budgets above, hence the band-edge tone sets the requirement.
To summarize, inter-slice amplitude matching is not the critical constraint; the clock arrival at the slices is (Section III-C), while intra-slice unit matching sets the in-band SNDR (Sections III-D and III-E).
To this point, the skew considered is only at the instant when a slice’s data changes, and Section III-C shows that it is not the dominant timing error.
Clock Consideration on the Timing Budget
Each slice is timed by two clocks: its phase of the slice update clock, at and delayed by , which delivers its new input code, and the clock that switches its current sources or capacitors. In the proposed implementation the baseband word is retimed by the clock at the slice, so the update instant follows the local edge and only has to land inside the retiming window. is the arrival skew of the clock at slice [36]. The slice output , with the carrier, has the baseband equivalent : besides delaying the data, the skew rotates the slice’s carrier by a constant phase across the hold duration. A carrier rotation is a complex slice gain, so the skirt weights it like , and (9) becomes
with the DFT of . A carrier skew therefore acts as a quadrature gain error of , and its budget is the gain budget divided by . Its weight exceeds that of the data-update term by , i.e., 12.5 dB at and 19.0 dB at . With retiming, and the two terms combine into the weight of a whole-slice time shift, worst at . The phase clocks , in contrast, cause no spur when their skew stays inside the retiming window , and all slices clear it with probability for
Beyond this timing window, a slice will latch the wrong cycle and the worst image rises to −40.0 dBc at and −45.4 dBc at .
Assumed in the proposed design (retiming at each slice’s from (15) with ).
Shown in Table II are the analytically obtained timing skew error budgets to achieve the 55 dBc SFDR target. Each agrees to within 1% with a -trial Monte Carlo of (7) in which each slice gain also includes the carrier phase . What is observed is that the slice update clocks are not the critical elements. Because each slice retimes its data on the local edge, in the median case, the and slice update clocks can be skewed by up to 28 and 23 ps rms before a slice latches the wrong cycle. The critical clock is the clock itself, because its skew also rotates the carrier of each slice. Its arrival at the slices must match to 0.79 ps () and 0.67 ps () rms at the median, and to ps for 95% yield. Only differential skew between the slices is consequential; common skew to all of the slices does not result in a spur. Since the carrier rotation is a fixed phase per slice, it could be measured and removed by a digital phase rotation of each stream, as in the background calibration of time-interleaved ADCs [37]. In that case, only the data-update error would remain, with the few-picosecond budgets of (10).
| Skew source | ||
| Data update, not retimed | 3.26 / 1.70 | 3.66 / 2.27 |
| carrier only | 0.96 / 0.48 | 0.66 / 0.33 |
| arrival, retimed* | 0.79 / 0.42 | 0.67 / 0.35 |
| Phase clock , retimed* | 27.7 / 15.7 | 22.5 / 14.3 |
Table II. RMS SKEW FOR 55 dBc SFDR, SINGLE TONE AT MHz, MEDIAN / 95% YIELD (PS)
In degrees of the slice update clock, a skew is . The phase-clock budgets of Table II are therefore () and () rms at the median, and and at 95% yield; without retiming they would be and . Since , the fixed slip window shrinks in degrees as grows, and the generator must be 1.8 to 2.5 times more accurate than the one.
Unit-Capacitor Mismatch
Section III-B treats each slice as a single weight with gain error . Inside a slice, however, the weight is built from unit capacitors, and their mismatch produces a code-dependent error. In the RF-DAC assumed for this design, each slice has a sign-magnitude I- and Q-array [38], each with thermometer-coded unit capacitors; the implemented macros drive units per array through their binary-to-thermometer decoders. Unit of array in slice has a fixed relative error , where the are independent and is the rms unit-capacitor mismatch [39].
Unit is switched on whenever the code magnitude is at least , and it then contributes according to the sign of the code. Its error therefore adds times this on/off pattern to the output, and the output spectrum at frequency bin is
where is the spectrum without mismatch, limited only by quantization, and is the spectrum of the on/off pattern of unit , held by slice as in (7). Because is a weighted sum of Gaussian errors, its real and imaginary parts are jointly Gaussian, centered on , with the covariance matrix . Rotating to the eigenvectors of , with eigenvalues and , splits the spur power into two independent terms,
where are the components of along the eigenvectors, divided by . This is a noncentral generalized distribution, and its cumulative distribution can be evaluated by a one-dimensional numerical integral.
The dominant spur is the image at caused by I/Q imbalance. If the I and Q arrays have slightly different total gains, a tone at leaks an image at whose amplitude is proportional to the gain difference, averaged over the slices. The phase of this image is fixed by the signal, and only its size and sign are random. One of the two terms in (17) therefore vanishes (), and the image power is the square of a single Gaussian variable, i.e., a scaled variable, as for the group in Section III-B. Averaging over the slices makes its variance fall as . Compared with the exponential distribution of a spur with random phase, the distribution has a long tail, which is why the 95%-yield budget below is about three times tighter than the median.
Spurious tones in different bins share the same unit errors; hence, they are correlated, but the distribution of the worst spur is bounded by , with the product as the point estimate. Only the 64 bins with the largest error variance or nominal level are kept; in a 2000-trial Monte Carlo simulation of the full spectrum the worst bin never fell outside of this set. For a 55 dBc SFDR the analysis gives a median budget of (, bounded by 10.2 and 10.8%) and 15.0% (, bounded by 14.8 and 15.3%) and a 95%-yield budget of 3.7% and 5.3%, for and , respectively. In this case, the bounds collapse because one spur is dominant. The median worst spur lies within 0.3 dB of the median I/Q image. Against the Monte Carlo, every point falls inside the 95% confidence interval of the simulated quantile, with a largest deviation of 0.4 dB.
In-Band SNDR Budget
The non-idealities in the prior subsections result in OOB spurs. However, with a modulated signal, the in-band SNDR is also important. The non-idealities are split into two classes. With static slice errors the array remains periodic in , so by (7) their errors land at , , outside the channel for , apart from a common gain and delay that any receiver removes. With these removed, a Monte Carlo simulation at the median budgets of Section III-B and Table II, with a 320 MHz orthogonal frequency-division multiplexing (OFDM)-like stimulus (10 dB peak-to-average power ratio (PAPR)), gives an in-band SNDR above 80 dB in 95% of the trials. Inter-slice mismatch is therefore an OOB emission problem only. Quantization, clock jitter and unit-capacitor mismatch, in contrast, place error inside the channel, and all of them add in power,
Each slice quantizes at , so a fraction of its noise falls in band, and recombination adds (Section III-A2). The resulting SNR is:
where is the Nyquist-band SNR of a single slice. The in-band quantization noise is thus set by , not by : for the aforementioned signal and resolution, simulated with the stimulus peak at 0.95 of full scale, dB and dB for both variants.
Clock Jitter
1) Clock Jitter: Unlike skew, jitter changes from edge-to-edge; hence, it raises the noise floor rather than producing a spur. As in Section III-C, the critical clock is . If its edge in RF cycle at slice is displaced by , the held input word of that slice is rotated by , and to a first order the output error is . The jitter is split into a common-mode (e.g., the source and the shared clock tree), with an rms value, and an independent mode due to each slice’s local buffer, with rms . The common part rotates the whole output, whereas the local parts add in power, and since for an oversampled signal they are averaged by . For white jitter, a fraction of the error falls in band, giving:
For dB this allows fs, or fs () and 396 fs (), and (20) agrees with a Monte Carlo simulation to within 0.16 dB. The common term defines the phase-noise requirement of any RF-DAC at ; when it is not white, its contribution is the phase noise integrated over the channel. At GHz, fs corresponds to a white phase noise of dBc/Hz, i.e., to dBc integrated over the 320 MHz channel. Only the local term is specific to the architecture, and it is relaxed by . The jitter is required to be at femtosecond levels, whereas the static skew of Table II may reach 0.7 ps, because static errors leave only a common gain and delay in band, while jitter is white and falls in band directly. Because the data are retimed, the same edge also moves the data transition, but this adds only of the carrier term, at most dB (). The jitter of the phase clocks adds to their skew in (15) and is otherwise absorbed by the retiming.
Unit-Capacitor Mismatch
2) Unit-Capacitor Mismatch: The unit-capacitor errors of Section III-D produce, besides the I/Q image, code-dependent distortion, and with a modulated signal both lie in band. Since (16) is exactly linear in the unit errors, the in-band error after projecting out the common complex gain with is a quadratic form,
where is the in-band signal power, collects the unit errors , and the columns of are their in-band spectra . As for the spurs, this is a weighted sum of squared Gaussian variables, and its CDF follows from Imhof’s integral [40]. Two properties set the budgets. First, the mean error power, , falls as , because the errors of different slices are independent and average out. Second, a single term dominates: the I/Q image carries 75% of the error power, so the in-band error inherits the long tail of the image spur, and the 95%-yield budget is much tighter than the median one. For dB, the unit capacitors must match to () and 7.5% () at the median, and to 2.4% and 3.4% at 95% yield. An independent time-domain Monte Carlo simulation of 2000 trials reproduces the mean to within 1% and both quantiles to within 6% in power. These budgets are 1.6 to 2 times tighter than the SFDR budgets of Section III-D: the SFDR counts only the largest spur, whereas the in-band error collects all of the distortion and, relative to the signal, grows with the back-off. In-band SNDR, not SFDR, therefore sets the size of the unit capacitors.
| Term | ||
| Quantization | 61.4 dB | 61.4 dB |
| Common jitter | 140 fs | 140 fs |
| Local jitter | 280 fs | 396 fs |
| for median / 95% yield | 5.3 / 2.4% | 7.5 / 3.4% |
| Inter-slice mismatch | dB | dB |
| Total SNDR, (18) | 51.5 dB | 51.5 dB |
Table 3. In-band SNDR budget, for the 320 MHz OFDM stimulus
Allowed value for 55 dB from that term alone. At the median SFDR budgets, 95% of trials; gain and delay over-rode.
Table III summarizes the in-band budget. Like the SFDR budgets, the jitter and unit-capacitor budgets scale by for a different target . With quantization at 61.4 dB and jitter and unit-capacitor mismatch each at 55 dB, (18) gives 51.5 dB.
Optimizing the Number of Phases, M
By using the yield laws derived in Section III along with the timing overheads of a target process and standard cell library, we can determine how every requirement scales with , before any implementation, by first establishing digital and analog design budgets and constraints.
Digital Lower Bound
The fastest digital clock runs at (§ 2-A), so each register stage has a cycle of , where . Out of this, every stage requires a fixed overhead that does not depend on : the flip-flop clock-to-Q and setup times, , and the jitter and skew of the clock. The time left for combinatorial logic is then:
which grows linearly with . Pipelining cannot recover a negative , because every added register requires the same overhead. Hence, once , a direct-digital implementation () cannot close timing at any pipeline depth. Even a positive is not sufficient: the datapath contains stages that cannot be split further, and the longest of them, which no further pipelining can shorten, sets a minimum logic delay . Requiring gives the lower bound
or, equivalently, for a fixed timing can be closed up to . The bound depends only on the technology/library for the implementation and the longest indivisible stage, not on the filter length, and on only through whether (commutator crossing) or (adder tree).
Cost of Raising on Analog/RF
The analog budgets of Section III depend on only through the ratio (i.e., the position of the band edge relative to the nulls from the zero-order hold at ). Hence, for a given signal bandwidth and , the dimensionless budgets (gain, unit-capacitor, and phase-clock skew in degrees) are functions of alone; the skew and jitter budgets additionally scale as . Increasing relaxes three of the specifications: the unit-capacitor budget grows as , because the I/Q image averages over independent slices, the local jitter budget also grows as , and the data-update skew budget increases, although retiming absorbs that skew in any case. The common jitter budget is independent of , and all other budgets tighten, as shown here for a 55 dBc SFDR at median yield with MHz and GHz ():
gain matching, from 7.7% at to 1.8% at , because the images move up the hold’s skirt as grows;
the arrival, from 0.79 to 0.23 ps;
the multiphase generator, from 32° to 2.4° rms, because its slip window is fixed in time while its period shrinks in degrees.
Earlier analysis focused on specific cases of . However, Monte Carlo simulations of other values of also support the yield laws. A -trial Monte Carlo at and , with each mismatch set to its predicted median budget, lands within 0.13 dB of the 55 dBc target. The digital energy and area also grow with the number of slices, as each slice carries its own decoder and output drivers.
The unit-capacitor relaxation should not be extrapolated far. The slice-differential spur sits 18 dB below the I/Q image at , 11 dB at , 6 dB at , and reaches it at . The image falls as while the differential spur grows with . For , the differential spur sets the unit-capacitor budget.
Design Rule
The analysis above shows that, while increasing relaxes the digital timing, it tightens most of the analog requirements. Hence, the best design choice is the smallest that satisfies (23), rounded up to the next divisor of (a power of two for ,
As an example, consider the 28-nm library used in Section V. In the slow-slow (SS) corner the flip-flop overhead is ps, and with budgets of ps and ps for clock jitter and skew, every stage requires 174 ps. The longest indivisible stage of the 4-phase datapath, the crossing from the FIR branches into the digital commutators, takes ps. At GHz, (23) then requires , so . This choice holds for GHz, and is required up to 29.4 GHz. For there are no digital commutators, and is set instead by the sub-filter adder tree (Section VI); the 29.4 GHz figure, like the column of Table IV, therefore assumes the combiner and is an estimate. Since the FR3 range targeted in Section II-B extends beyond 14.7 GHz, both variants are implemented in Section V: the 4-phase variant as the choice at 12.8 GHz, and the 8-phase variant as the candidate for the upper FR3 range (signed off here at 16 GSps) and to quantify the digital cost of raising at a fixed . Table IV summarizes the resulting requirements for to 32 at GHz.
| (ps) | (GHz) | (%) | (ps) | (°) | (%) | |
| 2 | −18 | 7.4 | 14.5 | 1.26 | 86 | 7.4 |
| 4 | 138 | 14.7 | 7.7 | 0.79 | 32 | 10.5 |
| 8 | 451 | 29.4 | 5.3 | 0.67 | 13 | 15.0 |
| 16 | 1076 | 58.9 | 3.4 | 0.44 | 5.5 | 21.2 |
| 32 | 2326 | 118 | 1.8 | 0.23 | 2.4 | n/a |
Table 4. Requirements Versus for 55 dBc SFDR at GHz. Median-yield budgets. : arrival skew, retimed. : phase-clock skew in slice update-clock degrees. Extrapolated as . Set by the slice-differential spur, not by scaling. assumed for . uses the 4-phase combiner ps; for , is set by the sub-filter adder tree and is lower (Section VI).
Circuit Implementation and Layout
To validate the physical viability of the proposed hybrid architectures, both macros were implemented up to the layout level: the 4-phase variant, the choice of (24) at 12.8 GHz, and the 8-phase variant, to quantify the cost of raising . The off-chip Farrow front-end resamples the modem stream to MSps, and the macro performs the high-speed integer-rate interpolation, to be paired with a custom switched-capacitor power amplifier (SCPA) acting as the RF-DAC for the passive analog recombination [22]. On the macro, an integer-rate FIR interpolator first raises the sample rate from to the interleaved rate GSps. Then, the polyphase FIR bank completes the interpolation. For the 8-phase variant, the data is streamed out of the macro at GSps, to be recombined on the RF-DAC. For the 4-phase variant, an additional digital commutator recombines the signals into 4-phases operating at GSps, to be recombined on the RF-DAC.
The digital polyphase FIR filter branches were generated as RTL from the Simulink model with HDL Coder and synthesized into a standard-cell gate-level netlist, optimizing the critical timing paths to ensure strict timing closure at 3.2 GHz () for the 4-phase variant and 1.6 GHz () for the 8-phase variant. The 8-phase macro was signed off at a 2 GHz slice update clock (16 GSps effective output rate), 25% above the nominal 1.6 GHz, with propagated clocks in the SS corner (0.81 V, 125°C), using low-threshold-voltage (LVT) standard cells only. The 4-phase macro was signed off at 3.6 GHz (14.4 GSps effective output rate), 12.5% above the nominal 3.2 GHz, in the same corner, using ultra-low-threshold-voltage (ULVT) standard cells only; its polyphase FIR branches run at half this rate. Figure 3 shows the placed-and-routed layouts of both digital macros.

Figure 3. Placed-and-routed layouts of the digital macros in 28-nm CMOS. (a) 4-phase macro, 776 × 652 µm (0.51 mm²). (b) 8-phase macro, 1024 × 716 µm (0.73 mm²).
The physical routing of this multi-phase clock network presents a primary design challenge, particularly for the 8-phase variant, which distributes eight 45°-spaced clock phases.
Because each slice retimes its data on , skew in this network causes no spur as long as it stays inside the slip window of (15), i.e., 22.5 ps rms at at the median (Table II). A symmetric, balanced H-tree routing topology was used across the layout, isolating the clock lines to minimize cross-coupling and differential phase skew [41]. The distribution, which sets the binding sub-picosecond requirement, resides in the RF-DAC and is outside the scope of these macros.
Post-Layout Validation and Results
Both macros were placed and routed in Cadence Innovus, and their parasitic resistances and capacitances (R+C) were extracted with Quantus. The implementation results of this section come from sign-off analysis of the extracted post-route netlists: static timing in Tempus, with propagated clocks, and power in Voltus, with switching activity from a gate-level Verilog simulation of the post-route netlist. To exercise the datapath with realistic activity, an 802.11be-like 320 MHz signal, chosen for its bandwidth and PAPR, was processed through the behavioral MATLAB/Simulink model of the Farrow filter, and the resulting bitstreams drove the gate-level simulation. The output bit patterns of the post-route netlist, ideally reconstructed, were identical to those of the Simulink model over the full 40,799-sample burst, confirming that both macros are bit-exact. This check is a zero-delay simulation and verifies function only; timing is covered by the static sign-off.
The impact of the non-idealities of Section III on the SFDR is evaluated with a behavioral model of the 4- and 8-phase recombiners, with slice timing and gain errors that model clock-routing asymmetries and unit-capacitor mismatch.
Figure 4 verifies the yield law of Section III-B: for each mechanism, the worst spur of trials of the exact, non-linearized (7), normalized to the largest group mean , follows (12) for both variants. For gain error, the and curves nearly coincide, because the group dominates. For timing skew, all groups have the same mean, so the curve lies to the right, which is the yield cost of a larger .

Figure 4. CDF of the worst mismatch spur for (a) timing skew and (b) gain error, for a single tone at MHz. The spur power is normalized to , the mean power of the strongest spur group across chips, from (10)–(11). Markers: exact Monte Carlo of (7) over trials; black lines: (12). Dotted lines mark the median and 95% yield.
Figure 5 shows the output spectrum for a modulated stimulus with both errors applied. With matched slices, only the signal band and the full-rate images at remain. Mismatch adds a copy of the band at every , about 60 dB below the in-band power spectral density (PSD) at its edges for the draw shown. Each copy straddles a hold null, so its center is notched and its band edges set its level, as assumed in Section III-B.

Figure 5. Simulated output spectrum for a 320 MHz flat, random-phase multitone at GHz for (a) and (b) , with slice timing skew ps and gain error ; the median-SFDR draw of 1001 Monte Carlo draws is shown. The PSD is normalized to the in-band PSD. Grey: hold envelope , with nulls at , . Each mismatch image is the signal band shifted by , so it is centered on a null that notches its middle; with matched slices these images vanish. The bands at are images of the recombined full-rate DAC, not mismatch products.
For unit-capacitor mismatch, the output error splits exactly into a common-slice part, which by (6) is a single full-rate DAC carrying the errors averaged over the slices, and a differential-slice part, the only one that can break the polyphase cancellation. The worst resulting spur is the I/Q image of the common part; odd harmonics lie about 7 dB lower, and the differential part stays below the image up to (Section IV). Averaging over the slices therefore gains about over a single array, and since standard digital I/Q calibration removes the image, the budgets of Section III-D are conservative [42].
Table V provides a quantitative post-layout performance breakdown contrasting the conventional direct-digital approach against the implemented 4-phase and 8-phase hybrid variants.
| Metric | Conventional Direct-Digital | 4-Phase Variant | 8-Phase Variant |
| Nominal Output Rate | 12.8 GSps | 12.8 GSps | 12.8 GSps |
| Sign-off Rate (SS) | n/a | 14.4 GSps | 16 GSps |
| Nominal Digital Clock | 12.8 GHz | 3.2 GHz | 1.6 GHz |
| RF-DAC Slices () | 1 | 4 | 8 |
| Clock Phase Spacing | Single Phase | Quadrature (90°) | Multi-Phase (45°) |
| Timing Closure (SS) | Not achievable by (23) | Met at 3.6 GHz ¶ | Met at 2 GHz § |
| Macro Energy / Sample (TT) ∥ | — | 27.2 pJ ¶ | 37.0 pJ § |
| Power TT / SS / FF (mW) ∥ | — | 391 / 414 / 459 | 592 / 558 / 710 |
| Clock Tree, TT | — | 30.0% | 29.2% |
| Leakage, TT / SS | — | 6.3 / 27.4% | 2.6 / 15.5% |
| Digital Macro Area | — | 0.51 mm¶ | 0.73 mm § |
| Routing Congestion | n/a | Medium | High |
Table V. Post-Layout Performance and Complexity Comparison
The benefit of a larger is timing feasibility rather than TT: 0.9 V, 25°C; SS: 0.81 V, 125°C; FF: 0.99 V, −40°C.
§ Implemented 8-phase macro, signed off and characterized at a 2 GHz slice update clock (16 GSps output).
¶ Implemented 4-phase macro, signed off and characterized at a 3.6 GHz slice update clock (14.4 GSps output).
∥ Whole digital macro, including the ×4 interpolator.
a reduction in arithmetic, as predicted by (23). The 4-phase macro closes at 3.6 GHz (14.4 GSps), 12.5% above nominal and just below the 14.7 GHz bound of Section IV; its critical path, the crossing from the half-rate FIR branches into the full-rate digital commutators, has 6 ps of slack. The 8-phase macro closes timing at 2 GHz, where its critical adder-tree stage meets setup with 17 ps of slack in the SS corner without additional pipelining.
Although the FIR arithmetic is the same for both variants, the implemented 8-phase macro carries about 1.9× the datapath cell area of the 4-phase macro, excluding the ×4 interpolator (1.4× in die area, 0.73 versus 0.51 mm): it drives twice as many RF-DAC slices, each with its own thermometer decoder, and it must span the larger analog core of the 8-phase array, with correspondingly longer interconnect from the FIR branches to the lanes. Its energy per output sample, each macro at its own sign-off rate in the TT corner, is nevertheless only 1.4× higher: 37.0 versus 27.2 pJ for the whole macro (Table V), or 34.5 versus 24.5 pJ for the FIR datapath alone, excluding the ×4 interpolator. It closes timing with LVT standard cells, whereas the 4-phase macro requires ULVT cells throughout. At their sign-off rates, the 8-phase variant thus buys timing margin for about 1.4× the energy per sample, consistent with the 4-phase choice of (24). A linear fit through the two macros gives about 17.4 + 2.45 pJ per sample and 0.29 + 0.056 mm; with two parameters for two points, mixed threshold-voltage flavors and sample rates, it indicates the trend rather than validating it.
Post-route power of both macros was annotated with switching activity from a zero-delay gate-level simulation of the post-route netlist driven by the 802.11be-like 320 MHz stimulus (100% of nets annotated; glitch power not included, so the reported energies are lower bounds), with each macro at its sign-off rate (Table V). In the TT corner, the ×4 interpolator consumes 40 mW (6.8%, 8-phase) and 37.5 mW (9.6%, 4-phase), and the clock tree about 30% in both, so both macros are dominated by dynamic power and benefit directly from the reduced slice update clock. Leakage rises to 27.4% of the 4-phase total at 125°C, because closing timing at 3.6 GHz required ULVT cells throughout.
Conclusion
This paper generalized polyphase interpolation for RF-DACs by splitting the recombination between digital commutators and the RF-DAC itself, which removes the full-rate output commutator and confines digital timing closure to . This is particularly useful for wideband RF-DACs, which cannot use high-order active reconstruction filters to suppress OOB emissions. The slice array is equivalent to a single full-rate DAC with an -times longer hold, and the yield laws, confirmed by Monte Carlo simulation, show that with retiming the binding requirement is sub-picosecond matching of the clock arrival across slices, and that the best choice is the smallest that closes digital timing.
Both implemented 28-nm macros close timing above their nominal rates in the SS corner: the 4-phase macro at 3.6 GHz (14.4 GSps) and the 8-phase macro at 2 GHz (16 GSps). The choice between them is a trade between timing feasibility and cost, not a power saving. The 8-phase variant consumes about 1.4× the energy per output sample and die area of the 4-phase variant and needs eight 45°-spaced clock phases, but closes timing with LVT cells and has lower leakage at 125°C. The 4-phase variant therefore suits platforms that favor energy and only four clock phases, and the 8-phase variant suits those that need timing margin, high-temperature operation, or headroom to raise the output rate.
References
- [1]J. Mitola, “The software radio architecture,” IEEE Commun. Mag., vol. 33, pp. 26–38, May 1995.DOI
- [2]A. O. Watanabe, M. Ali, S. Y. B. Sayeed, R. R. Tummala, and M. R. Pulugurtha, “A review of 5g front-end systems package integration,” IEEE Transactions on Components, Packaging and Manufacturing Technology, vol. 11, no. 1, pp. 118–133, 2021.
- [3]S. Luschs, R. Schreier, and H.-S. Lee, “Radio frequency digital-to-analog converter,” IEEE J. Solid-State Circuits, vol. 39, pp. 1462–1467, Sep. 2004.
- [4]S.-M. Yoo, J. S. Walling, E. C. Woo, B. Jann, and D. J. Allstot, “A switched-capacitor RF power amplifier,” IEEE J. Solid-State Circuits, vol. 46, pp. 2977–2987, Dec. 2011.
- [5]M. S. Alavi, R. B. Staszewski, L. C. N. de Vreede, and J. R. Long, “A wideband 2× 13-bit all-digital I/Q RF-DAC,” IEEE Trans. Microw. Theory Tech., vol. 62, pp. 732–752, Apr. 2014.DOI
- [6]D. Shakya et al., “Comprehensive rf(c) and fr3 lower and upper mid-band propagation and material penetration loss measurements and channel models in indoor environment for 5g and 6g,” IEEE Open Journal of the Communications Society, vol. 5, pp. 5192–5218, 2024.
- [7]A. Bhide and A. Alvandpour, “An 11-GS/s 1.1-GHz bandwidth interleaved ΔΣ DAC for 60-GHz radio in 65-nm CMOS,” IEEE J. Solid-State Circuits, vol. 50, pp. 2306–2318, Oct. 2015.DOI
- [8]E. Oljemann, A.-J. Annema, and B. Nauta, “An interleaved full Nyquist high-speed DAC technique,” IEEE J. Solid-State Circuits, vol. 50, pp. 704–713, Mar. 2015.
- [9]P. Caragiulo, O. E. Mattia, A. Arbabian, and B. Murmann, “A 2× time-interleaved 28-GS/s 8-bit 0.03-mm² switched-capacitor DAC in 16-nm FinFET CMOS,” IEEE J. Solid-State Circuits, vol. 56, pp. 2335–2346, Aug. 2021.DOI
- [10]J. J. McCue et al., “A time-interleaved multimode ΔΣ RF-DAC for direct digital-to-RF synthesis,” IEEE J. Solid-State Circuits, vol. 51, pp. 1109–1124, May 2016.DOI
- [11]D. Su and B. Wooley, “A CMOS oversampling D/A converter with a current-mode semidigital reconstruction filter,” IEEE J. Solid-State Circuits, vol. 28, pp. 1224–1233, Dec. 1993.DOI
- [12]D. Barkin, A. Lin, D. Su, and B. Wooley, “A CMOS oversampling bandpass cascaded D/A converter with digital FIR and current-mode semi-digital filtering,” IEEE J. Solid-State Circuits, vol. 39, pp. 585–593, Apr. 2004.
- [13]S. Taleie, T. Copani, B. Bakkaloglu, and S. Kiaei, “A bandpass ΔΣ RF-DAC with embedded FIR reconstruction filter,” in IEEE ISSCC Dig. Tech. Papers, 2006, pp. 2370–2379.DOI
- [14]R. Bhat, J. Zhou, and H. Krishnaswamy, “Wideband mixed-domain multi-tap finite-impulse response filtering of out-of-band noise floor in watt-class digital transmitters,” IEEE J. Solid-State Circuits, vol. 52, pp. 3405–3420, Dec. 2017.DOI
- [15]S. Su and M. S.-W. Chen, “A SAW-less direct-digital RF modulator with tri-level time-approximation filter and reconfigurable dual-band delta-sigma modulation,” in IEEE ISSCC Dig. Tech. Papers, 2020, pp. 174–176.
- [16]S. Su and M. S.-W. Chen, “A time-approximation filter for direct RF transmitter,” IEEE J. Solid-State Circuits, vol. 56, pp. 2018–2028, Jul. 2021.DOI
- [17]W. Black and D. Hodges, “Time interleaved converter arrays,” IEEE J. Solid-State Circuits, vol. 15, pp. 1022–1029, Dec. 1980.
- [18]J. Muller et al., “A 7-bit 18th order 9.6 GS/s FIR up-sampling filter for high data rate 60-GHz wireless transmitters,” IEEE J. Solid-State Circuits, vol. 47, pp. 1743–1756, Jul. 2012.DOI
- [19]C. Farrow, “A continuously variable digital delay element,” in Proc. IEEE Int. Symp. Circuits Syst. (ISCAS), 1988, pp. 2641–2645.
- [20]T. Hentschel and G. Fettweis, “Sample rate conversion for software radio,” IEEE Commun. Mag., vol. 38, pp. 142–150, Aug. 2000.DOI
- [21]M. Hashemi, Y. Shen, M. Mehrpoo, M. S. Alavi, and L. C. N. de Vreede, “An intrinsically linear wideband polar digital power amplifier,” IEEE J. Solid-State Circuits, vol. 52, pp. 3312–3328, Dec. 2017.DOI
- [22]W. Yuan and J. S. Walling, “A multiphase switched capacitor power amplifier,” IEEE J. Solid-State Circuits, vol. 52, pp. 1320–1330, May 2017.DOI
- [23]R. E. Crochiere and L. R. Rabiner, Multirate Digital Signal Processing. Englewood Cliffs, NJ, USA: Prentice-Hall, 1983.DOI
- [24]P. P. Vaidyanathan, Multirate Systems and Filter Banks. Englewood Cliffs, NJ, USA: Prentice-Hall, 1993.
- [25]F. J. Harris, Multirate Signal Processing for Communication Systems. Upper Saddle River, NJ, USA: Prentice-Hall, 2004.
- [26]R. Khoini-Poorfard, L. Lim, and D. Johns, “Time-interleaved oversampling A/D converters: Theory and practice,” IEEE Trans. Circuits Syst. II, vol. 44, pp. 634–645, Aug. 1997.
- [27]Y.-C. Tsao and K. Choi, “Area-efficient parallel FIR digital filter structures for symmetric convolutions based on fast FIR algorithm,” IEEE Trans. VLSI Syst., vol. 20, pp. 366–371, Feb. 2012.DOI
- [28]E. Hogenauer, “An economical class of digital filters for decimation and interpolation,” IEEE Trans. Acoust., Speech, Signal Process., vol. 29, pp. 155–162, Apr. 1981.DOI
- [29]A. Kwentus, Z. Jiang, and A. Willson, “Application of filter sharpening to cascaded integrator-comb decimation filters,” IEEE Trans. Signal Process., vol. 45, pp. 457–467, Feb. 1997.DOI
- [30]S. Balasubramanian et al., “Systematic analysis of interleaved digital-to-analog converters,” IEEE Trans. Circuits Syst. II, vol. 59, pp. 83–87, Feb. 2012.DOI
- [31]L. Duncan et al., “A 10-bit DC-20-GHz Multiple-Return-to-Zero DAC With >48-dB SFDR,” IEEE J. Solid-State Circuits, vol. 52, pp. 3262–3275, Dec. 2017.DOI
- [32]Y.-C. Jenq, “Digital spectra of nonuniformly sampled signals: Fundamental and high-speed waveform digitizers,” IEEE Trans. Instrum. Meas., vol. 37, pp. 245–251, Jun. 1988.DOI
- [33]N. Kurosawa, H. Kobayashi, K. Maruyama, H. Sugawara, and K. Kobayashi, “Explicit analysis of channel mismatch effects in time-interleaved ADC systems,” IEEE Trans. Circuits Syst. I, vol. 48, pp. 261–271, Mar. 2001.DOI
- [34]M. El-Chammas and B. Murmann, “General analysis on the impact of phase-skew in time-interleaved ADCs,” IEEE Trans. Circuits Syst. I, vol. 56, pp. 902–910, May 2009.
- [35]G. Leger, E. Peralias, A. Rueda, and J. Huertas, “Impact of random channel mismatch on the SNR and SFDR of time-interleaved ADCs,” IEEE Trans. Circuits Syst. I, vol. 51, pp. 140–150, Jan. 2004.DOI
- [36]B. Razavi, “Design considerations for interleaved ADCs,” IEEE J. Solid-State Circuits, vol. 48, pp. 1806–1817, Aug. 2013.DOI
- [37]J. Elbornsson, F. Gustafsson, and J.-E. Eklund, “Blind adaptive equalization of mismatch errors in a time-interleaved A/D converter system,” IEEE Trans. Circuits Syst. I, vol. 51, pp. 151–158, Jan. 2004.DOI
- [38]M. Mehrpoo, M. Hashemi, Y. Shen, L. C. N. de Vreede, and M. S. Alavi, “A wideband linear I/Q-interleaving DDRM,” IEEE J. Solid-State Circuits, vol. 53, pp. 1361–1373, May 2018.DOI
- [39]J.-B. Shyu, G. Temes, and F. Krummenacher, “Random error effects in matched MOS capacitors and current sources,” IEEE J. Solid-State Circuits, vol. 19, pp. 948–956, Dec. 1984.DOI
- [40]J. P. Imhof, “Computing the distribution of quadratic forms in normal variables,” Biometrika, vol. 48, no. 3/4, pp. 419–426, 1961.DOI
- [41]E. Friedman, “Clock distribution networks in synchronous digital integrated circuits,” Proc. IEEE, vol. 89, pp. 665–692, May 2001.DOI
- [42]L. Anttila, M. Valkama, and M. Renfors, “Frequency-selective I/Q mismatch calibration of wideband direct-conversion transmitters,” IEEE Trans. Circuits Syst. II, vol. 55, pp. 359–363, Apr. 2008.DOI