Pith. sign in

REVIEW 3 major objections 4 minor 17 references

This paper claims that a SAR ADC can replace its first four comparator-driven decisions with a Kalman-filter prediction of the top bits, cutting total power by 50.3% at 20 MS/s.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-01 21:13 UTC pith:F32BND4B

load-bearing objection Plausible predictive-SAR idea, but the 'measured' silicon claims are contradicted by the paper's own simulation-only sections, and the power reduction doesn't include the Kalman filter's cost. the 3 major comments →

arxiv 2607.16139 v1 pith:F32BND4B submitted 2026-07-17 eess.SP cs.SYeess.SY

A Kalman Filter-Assisted Data-Predictive SAR ADC With Reduced Switching Energy for Low-Power Applications

Xiyuan Feng , Yuxiang Zhao , Jie Xiong , Dian Lin , Yunlei Zhong , Wei Liu , Zhongheng Ji , Ruiyu Tian
show 2 more authors
Chenhao Zhuo Yue Yin
This is my paper
classification eess.SP cs.SYeess.SY
keywords SAR ADCKalman filterdata predictionMSB predictionswitching energylow-power ADCdual-mode conversion180-nm CMOS
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper tries to establish that the conventional N-cycle, bit-by-bit SAR conversion flow is not mandatory: a Kalman filter running in the background can predict the four most significant bits of the next sample from the complete conversion history, so those bits' capacitors can be switched in parallel and the first four comparator cycles skipped entirely. If true, this breaks the fixed relationship between resolution and conversion energy/latency for a whole class of slowly varying signals (biomedical, sensor). The reported payoff is a 50.3% total power reduction at 20 MS/s (1.96 mW down to 0.975 mW) with SNR/SFDR of 57.88/74.51 dB at 504 kHz, plus a dual-mode fallback that preserves accuracy during transients. The central evidence is a circuit-level design in 180-nm CMOS with an algorithm-level Kalman model; the digital filter's hardware implementation and co-simulation are left as future work. A sympathetic reader would take the paper as a feasibility demonstration that predictive digital estimation can be the primary energy saver in a SAR ADC.

Core claim

The ADC's defining claim is that the first four comparator-driven SAR decisions can be replaced by a two-state Kalman predictor—amplitude plus local slope—so the four MSB capacitors are preset in parallel before the comparator strokes. The completed 10-bit code feeds back as the measurement, and because the ADC output is already the digital representation, the amplitude state is reset directly to that code while only the slope state is filtered. This predict–quantize–update loop skips four clock cycles, and an optimized 4-bit switching table selects the minimum number of capacitor flips per code, cutting switching events by about half. The paper reports a 50.3% total power reduction in predi

What carries the argument

The load-bearing element is a second-order Kalman predictor with a constant-velocity state model: the state vector holds the signal code and its per-sample slope; the prediction step is a single multiply-add; and the update step resets the amplitude directly to the measured code while adapting the slope through a scalar gain α. Surrounding it, the optimized 4-bit MSB switching scheme precomputes the minimal capacitor transitions for each of the 16 top-bit codes, and a dual-mode controller (two-consecutive-saturation hysteresis) falls back to full 10-cycle conversion when the residual leaves the 6-bit correction window. The whole argument rests on the claim that four MSB comparisons can be tr

Load-bearing premise

The power-saving claim collapses if the digital Kalman predictor cannot be implemented at 20 MS/s with negligible area, power, and timing overhead—the paper models the filter at algorithm level and defers RTL synthesis and co-simulation to future work.

What would settle it

Run a post-layout co-simulation of the synthesized Kalman filter and ADC core at 20 MS/s and measure total power in predictive mode; if the digital block's power plus any timing-induced settling margin pushes total power above 0.975 mW, or if the predictive-mode FFT cannot be reproduced with the filter physically in the loop, the central claim is refuted.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • Predictive mode shortens the conversion from ten cycles to six-plus-prediction, so latency per sample drops by four clock periods for correlated inputs.
  • The optimized 4-bit switching table cuts MSB capacitor transition events by about half across all output codes, reducing CDAC dynamic energy.
  • At 20 MS/s and 1.8 V, total power falls from 1.96 mW to 0.975 mW (50.3%), yielding an FoM of 123.1 fJ/conv-step in predictive mode.
  • The dual-mode fallback means the predictor can be enabled on real data streams without risking corrupted samples during transients.
  • Scaling the same architecture to 40/28-nm nodes is projected to bring power into the microwatt range with 30–40% better energy efficiency.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The energy saving is signal-statistics-dependent: the 50.3% figure holds for inputs that stay within the 6-bit correction window about 94% of the time (as modeled on ECG/EEG), so sine waves at full scale or broadband inputs would yield a smaller gain.
  • A simpler first-order extrapolation might capture a large share of the same savings; the Kalman filter's added value over a cheaper predictor would need a direct accuracy-and-power comparison on identical inputs to justify its overhead.
  • The predict-and-switch-in-parallel idea is not tied to Kalman filtering specifically; it could be combined with redundancy-based error correction or applied to predict more than four bits, trading prediction coverage against savings.
  • The decisive unknown is the digital implementation: the paper's algorithm-level model must be turned into synthesizable RTL and co-simulated, and that co-simulation may shrink the net power benefit once filter logic, clocking, and routing are counted.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 4 minor

Summary. The paper proposes a 10-bit 180-nm SAR ADC in which a Kalman-filter-based recursive estimator predicts the four most significant bits before each conversion, allowing these MSB capacitors to be switched simultaneously and four comparator cycles to be skipped. The authors claim a 50.3% total power reduction (1.96 mW to 0.975 mW) at 20 MS/s under a 1.8 V supply, a reduced quantization cycle, and dynamic performance of SNR/SFDR = 57.88/74.51 dB. The architecture includes a dual-mode fallback for prediction failures and an optimized 4-bit MSB switching scheme. The central claim, however, is not supported by the evidence presented: Section V is a simulation section, Section VI-A states that the Kalman filter is only behaviorally modeled and that RTL co-simulation/synthesis is future work, and Section VII nevertheless claims a fabricated chip and 'measured' results. The abstract and conclusion also report conflicting dynamic performance numbers.

Significance. If the claimed results were actually obtained from a fabricated chip with a working Kalman-filter predictor, the work would be a meaningful contribution: skipping the first four SAR cycles with a low-cost digital predictor and halving total power at 20 MS/s would be of clear interest to low-power ADC designers. The algorithmic idea is clearly described and the optimized switching table is a useful starting point. However, because the key enabling component—the Kalman predictor—is explicitly not implemented in hardware or even RTL, the headline energy reduction cannot be verified. The paper's value as a simulation study is also diminished by the absence of the digital overhead estimate and by the internal contradictions in the reported dynamic-performance numbers.

major comments (3)
  1. [Sections V, VI-A, VII; Abstract] The central claim of a 50.3% total power reduction is unsupported. Section V is titled 'Simulation Results' and its opening paragraph says the ADC was evaluated through behavioral and circuit-level simulations, yet Fig. 7 and Fig. 8 are captioned 'measured.' Section VI-A explicitly states that the Kalman filter 'was modeled at the algorithm level' and that a dedicated hardware implementation, RTL co-simulation, and synthesis are future work. The conclusion nevertheless asserts the ADC was 'Fabricated in a 180-nm CMOS process' and reports measured power reduction. Since the predictive-mode power number cannot include the Kalman filter's own digital power, area, or timing overhead, the headline reduction from 1.96 mW to 0.975 mW is not established. This is a load-bearing contradiction, not a stylistic issue.
  2. [Abstract, Section V-B, Section VII] The dynamic performance numbers are internally inconsistent. The abstract and Section V-B report a measured SNR/SFDR of 57.88/74.51 dB at 504 kHz, while Section VII reports an SNDR/SFDR of 53.63/65.89 dB. These cannot both be the same measurement: the SFDR values differ by 8.6 dB, and one entry is SNR while the other is SNDR. Table II lists an ENOB of 8.63, which is consistent with the SNDR value but not with the SNR value. The authors must clarify which numbers are the actual simulated/measured results and explain the discrepancy; as written, the paper contradicts itself on a key performance claim.
  3. [Sections III-A.1, III-A.3, IV-A] The 94% in-window prediction rate and the 50.3% energy saving depend on empirically tuned parameters K, alpha, and M. The paper states that K=4 was chosen 'through MATLAB modeling across ECG and EEG waveforms' and that alpha was 'tuned empirically via MATLAB simulation to minimize the RMS prediction error over the target signal ensemble,' but no dataset, signal statistics, tuning protocol, or cross-validation is provided. Without this, the 94% success rate is an unspecified overfit to an undisclosed ensemble, and the fallback energy cost during the M-sample recovery period cannot be assessed. Given that the energy claim relies on a high successful-prediction rate, this missing information is load-bearing.
minor comments (4)
  1. [Section III-A.3, Eqs. (6)-(7)] The estimator in Eqs. (6)-(7) is an alpha-beta filter, not a general Kalman filter: the position state is set to the measurement and the velocity state is updated with a fixed gain alpha, with no noise covariance propagation or Riccati equation. The paper should either present the steady-state Kalman derivation that reduces to this form or refer to the estimator as an alpha-beta filter / constant-gain recursive predictor.
  2. [Section VI-C] The paper itself acknowledges in Section VI-C that 'comprehensive co-simulation with a synthesized RTL implementation of the filter and the ADC core is required to fully validate the system-level power and timing.' This limitation should be reflected in the abstract and conclusion, which currently present the results as measured silicon data.
  3. [Table II] The performance comparison in Table II mixes designs at very different sampling rates (0.15-0.25 MS/s versus 20 MS/s). The authors note this in Section VI-C, but the FoM comparison is still shown in the same table without normalization or a clear caveat.
  4. [General] Several figure captions use 'measured' for results that are simulations. Please change all such captions to 'simulated' unless actual silicon measurements are provided.

Circularity Check

1 steps flagged

The 94% in-window prediction rate is produced by tuning α and K on the same ECG/EEG ensemble, then quoted as a prediction result; the central power-saving architecture is otherwise not circular but is unverified at the digital-implementation level.

specific steps
  1. fitted input called prediction [Section III-A, Eqs. (6)-(7) and the 'Prediction Principle' paragraph]
    "Through MATLAB modeling across ECG and EEG waveforms, a 4-bit prediction was found to provide the optimal trade-off: it eliminates 4 of 10 comparison cycles while confining the residual δ(k) within the 6-bit correction window for over 94% of samples. ... α is tuned empirically via MATLAB simulation to minimize the RMS prediction error over the target signal ensemble."

    The success metric (94% of samples within the 6-bit window) is reported as justification that K=4 is sufficient, but α in Eq. (7) and K=4 were both selected by fitting to the same ECG/EEG waveform ensemble. The 94% figure is therefore the training-set performance of the tuned predictor, not an out-of-sample prediction. The paper then cites this fitted value as evidence that the predictive mode is robust and energy-saving, so the claimed prediction performance reduces to the fit. The CDAC switching-energy comparison itself is a direct enumeration and is not circular; the circularity is localized to the presented validation metric.

full rationale

The Kalman recursion is not self-definitional: the prediction x̂(k|k−1) uses only samples up to k−1, and Eq. (6) resets the amplitude state to the completed output z(k), so the current sample is not used to predict itself. The optimized 4-bit switching table is an explicit enumeration of capacitor configurations, not a hidden identity. The main non-circularity concern is that α and the 4-bit skip depth are empirically fitted on the target signal ensemble, and the quoted 94% in-window rate is an in-sample fit statistic. Separately, the paper labels Section V-B results as 'measured' while Section VI-A states the Kalman filter was modeled only at algorithm level with RTL co-simulation as future work; this is a verification gap and an internal inconsistency, but it is not a circularity and is not scored as one.

Axiom & Free-Parameter Ledger

3 free parameters · 4 axioms · 0 invented entities

The central claims rest on an implicit signal model (slowly varying, locally linear), an empirically tuned gain alpha, a chosen prediction depth K, and the unverified assumption that digital prediction overhead is negligible. No new physical entities are introduced.

free parameters (3)
  • alpha (slope update gain) = not stated
    Used in Eq. (7); tuned empirically via MATLAB to minimize RMS prediction error over the target signal ensemble. No value or sensitivity analysis is provided.
  • K (number of predicted MSBs) = 4
    Chosen via MATLAB modeling across ECG/EEG signals as the optimal trade-off. This design choice directly determines the claimed 4-cycle savings and 64-LSB correction window.
  • M (fallback recovery length) = not stated
    Number of samples spent in conventional mode after two consecutive saturation flags; said to be shorter than a QRS complex but never quantified.
axioms (4)
  • domain assumption Constant-velocity state model (Eqs. 2-3) captures the input signal between samples.
    The prediction is built on locally linear signal behavior. It is acknowledged to fail during transients such as QRS complexes.
  • domain assumption The ADC output is the true input measurement, so the amplitude state can be reset directly to z(k) with no sensor-noise model.
    Eq. (6) resets the state to the measured code; this simplifies the filter to an alpha tracker and ignores real input-referred noise.
  • ad hoc to paper ECG/EEG waveforms are representative of the target applications, and the 94% in-window rate measured on them supports the energy claim.
    K and alpha are chosen from this same signal ensemble, with no held-out validation or comparison to other input classes.
  • domain assumption The CDAC switching-energy and total-power simulation models are accurate enough to support a 50.3% reduction claim.
    No corner, Monte Carlo, mismatch, or measurement-uncertainty analysis is reported for the power numbers.

pith-pipeline@v1.3.0-alltime-deepseek · 10369 in / 15858 out tokens · 126597 ms · 2026-08-01T21:13:41.058237+00:00 · methodology

0 comments
read the original abstract

The proliferation of Internet of Things (IoT) devices and wearable health monitors has created an urgent demand for ultra-low-power analog-to-digital converters (ADCs). Successive approximation register (SAR) ADCs are widely used in such applications, yet their energy efficiency remains constrained by the sequential bit-by-bit switching of the capacitive DAC (CDAC). The high-weight most significant bit (MSB) transitions dominate the total switching energy, and the rigid N -cycle conversion flow imposes a hard lower bound on latency per sample.This paper presents a Kalman filter-assisted data-predictive SAR ADC that replaces the first four comparator-driven decisions with a recursive state estimator. The Kalman filter predicts the 4 MSBs from the complete conversion history before each cycle begins, enabling simultaneous parallel switching of the MSB capacitors. This eliminates redundant CDAC transitions, shortens the quantization cycle by four clock periods, and reduces switching energy by approximately 50%. An optimized 4-bit MSB switching scheme further suppresses residual switching at the hardware level. The ADC, designed in a 180-nm CMOS process, supports configurable dual-mode operation, toggling between a conventional mode and the Kalman-driven predictive mode for robustness under erratic inputs. At 20 MS/s and a 1.8-V supply, the predictive mode reduces total power consumption by 50.3% (from 1.96 mW to 0.975 mW), with a measured SNR/SFDR of 57.88/74.51 dB at 504 kHz, confirming its suitability for energy-constrained wireless sensor networks.

Figures

Figures reproduced from arXiv: 2607.16139 by Chenhao Zhuo, Dian Lin, Jie Xiong, Ruiyu Tian, Wei Liu, Xiyuan Feng, Yue Yin, Yunlei Zhong, Yuxiang Zhao, Zhongheng Ji.

Figure 1
Figure 1. Figure 1: Block diagram of the proposed SAR ADC integrated with a Kalman filter prediction engine. [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Simulated ADC quantization results: comparison between traditional [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Quantization error before and after Kalman filter integration. The [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Proposed SAR logic with integrated Kalman filter prediction module and optimized 4-bit MSB switching controller. The MUX2 block selects between [PITH_FULL_IMAGE:figures/full_fig_p006_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: DAC switch control logic. Each MSB switch is driven by a MUX2 [PITH_FULL_IMAGE:figures/full_fig_p006_5.png] view at source ↗
Figure 8
Figure 8. Figure 8: Output FFT spectrum of the proposed ADC sampling a 504 kHz [PITH_FULL_IMAGE:figures/full_fig_p007_8.png] view at source ↗
Figure 7
Figure 7. Figure 7: Measured total power consumption versus sampling frequency for the [PITH_FULL_IMAGE:figures/full_fig_p007_7.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

17 extracted references

  1. [1]

    A 0.3 V 10-bit 1.17 f SAR ADC with merge and split switching in 90 nm CMOS,

    J.-Y . Lin and C.-C. Hsieh, “A 0.3 V 10-bit 1.17 f SAR ADC with merge and split switching in 90 nm CMOS,”IEEE Trans. Circuits Syst. I: Reg. Papers, vol. 62, no. 1, pp. 70–79, 2015

    2015

  2. [2]

    A 10-bit 100-MS/s reference-free SAR ADC in 90 nm CMOS,

    Y . Zhu, C.-H. Chan, U.-F. Chio, S.-W. Sin, S.-P. U, R. P. Martins, and F. Maloberti, “A 10-bit 100-MS/s reference-free SAR ADC in 90 nm CMOS,”IEEE J. Solid-State Circuits, vol. 45, no. 6, pp. 1111–1121, 2010

    2010

  3. [3]

    A 0.6-V 38-nW 9.4-ENOB 20-kS/s SAR ADC in 0.18-µm CMOS for medical implant devices,

    Z. Zhu and Y . Liang, “A 0.6-V 38-nW 9.4-ENOB 20-kS/s SAR ADC in 0.18-µm CMOS for medical implant devices,”IEEE Trans. Circuits Syst. I: Reg. Papers, vol. 62, no. 9, pp. 2167–2176, 2015

    2015

  4. [4]

    A 1-µW 10-bit 200-kS/s SAR ADC with a bypass window for biomedical applications,

    G.-Y . Huang, S.-J. Chang, C.-C. Liu, and Y .-Z. Lin, “A 1-µW 10-bit 200-kS/s SAR ADC with a bypass window for biomedical applications,” IEEE J. Solid-State Circuits, vol. 47, no. 11, pp. 2783–2795, 2012

    2012

  5. [5]

    A 10-b ternary SAR ADC with quantization time information utilization,

    J. Guerber, H. Venkatram, M. Gande, A. Waters, and U.-K. Moon, “A 10-b ternary SAR ADC with quantization time information utilization,” IEEE J. Solid-State Circuits, vol. 47, no. 11, pp. 2604–2613, 2012

    2012

  6. [6]

    A 0.4 V 1.94 fJ/conversion-step 10 bit 750 kS/s SAR ADC with input-range-adaptive switching,

    P.-C. Lee, J.-Y . Lin, and C.-C. Hsieh, “A 0.4 V 1.94 fJ/conversion-step 10 bit 750 kS/s SAR ADC with input-range-adaptive switching,”IEEE Trans. Circuits Syst. I: Reg. Papers, vol. 63, no. 12, pp. 2149–2157, 2016

    2016

  7. [7]

    A 0.85 fJ/conversion- step 10b 200 kS/s subranging SAR ADC in 40 nm CMOS,

    H.-Y . Tai, Y .-S. Hu, H.-W. Chen, and H.-S. Chen, “A 0.85 fJ/conversion- step 10b 200 kS/s subranging SAR ADC in 40 nm CMOS,” in2014 IEEE ISSCC Dig. Tech. Papers, 2014, pp. 196–197

    2014

  8. [8]

    A bypass-switching SAR ADC with a dynamic proximity comparator for biomedical applications,

    T.-Y . Wang, H.-Y . Li, Z.-Y . Ma, Y .-J. Huang, and S.-Y . Peng, “A bypass-switching SAR ADC with a dynamic proximity comparator for biomedical applications,”IEEE J. Solid-State Circuits, vol. 53, no. 6, pp. 1743–1754, 2018

    2018

  9. [9]

    A 0.3 V 10-bit SAR ADC with first 2- bit guess in 90-nm CMOS,

    J.-Y . Lin and C.-C. Hsieh, “A 0.3 V 10-bit SAR ADC with first 2- bit guess in 90-nm CMOS,”IEEE Trans. Circuits Syst. I: Reg. Papers, vol. 64, no. 3, pp. 562–572, 2017

    2017

  10. [10]

    A 12-bit 40-MS/s SAR ADC with a fast-binary-window DAC switching scheme,

    Y .-H. Chung, C.-W. Yen, P.-K. Tsai, and B.-W. Chen, “A 12-bit 40-MS/s SAR ADC with a fast-binary-window DAC switching scheme,”IEEE Trans. VLSI Syst., vol. 26, no. 10, pp. 1989–1998, 2018

    1989

  11. [11]

    A 12-bit SAR ADC with a DAC-configurable window switching scheme,

    Y .-H. Chung, Q.-F. Zeng, and Y .-S. Lin, “A 12-bit SAR ADC with a DAC-configurable window switching scheme,”IEEE Trans. Circuits Syst. I: Reg. Papers, vol. 67, no. 2, pp. 358–368, 2020

    2020

  12. [12]

    Predicting adc: A new approach for low power adc design,

    N. Wood and N. Sun, “Predicting adc: A new approach for low power adc design,” in2014 IEEE Dallas Circuits and Systems Conference (DCAS), 2014, pp. 1–4

    2014

  13. [13]

    Time assisted SAR ADC with bit-guess and digital error correction,

    B. Canal, H. D. Klimach, S. Bampi, and T. R. Balen, “Time assisted SAR ADC with bit-guess and digital error correction,” inProc. 35th SBC/SBMicro/IEEE/ACM Symp. Integr. Circuits Syst. Design (SBCCI), 2022, pp. 1–6

    2022

  14. [14]

    Architectural exploration for energy-efficient fixed-point Kalman filter VLSI design,

    P. T. L. Pereira, G. Paim, P. A. L. da Costa, E. A. C. da Costa, S. J. M. de Almeida, and S. Bampi, “Architectural exploration for energy-efficient fixed-point Kalman filter VLSI design,”IEEE Trans. VLSI Syst., vol. 29, no. 7, pp. 1402–1415, 2021

    2021

  15. [15]

    Parameterizable FPGA-based Kalman filter coprocessor using piecewise affine modeling,

    A. Mills, P. H. Jones, and J. Zambreno, “Parameterizable FPGA-based Kalman filter coprocessor using piecewise affine modeling,” inProc. IEEE Int. Parallel Distrib. Process. Symp. Workshops (IPDPSW), 2016, pp. 139–147

    2016

  16. [16]

    A 10 bit SAR ADC with data- dependent energy reduction using LSB-first successive approximation,

    F. M. Yaul and A. P. Chandrakasan, “A 10 bit SAR ADC with data- dependent energy reduction using LSB-first successive approximation,” IEEE J. Solid-State Circuits, vol. 49, no. 12, pp. 2825–2834, 2014

    2014

  17. [17]

    A 510 nW 12-bit 200 kS/s SAR-assisted SAR ADC using a re-switching technique,

    Y .-S. Hu, H.-S. Chen, and K.-Y . Lin, “A 510 nW 12-bit 200 kS/s SAR-assisted SAR ADC using a re-switching technique,” inProc. IEEE ISCAS, 2017, pp. 1–4. JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2021 9 PLACE PHOTO HERE Michael ShellUse\begin{IEEEbiography} and then for the 1st argument use \includegraphicsto declare and link the author photo. ...

    2017