{"id":"a0276fac-6188-4b69-aa8c-1ea9a4e3a1d7","arxiv_id":"1908.01343","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":3.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"An exhaustive simulation of 20 approximate 8-bit multipliers shows that error distance, normalized error distance, and PSNR vary strongly with the input operands.","lead":"This report exhaustively tests 20 approximate 8-bit multiplier designs across all input pairs and shows that the error depends strongly on which numbers are multiplied. The finding could help error-tolerant systems avoid input ranges that produce unacceptably large errors.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Table 2 metrics contradict Eqs. (3) and (7), so the NED/PSNR surfaces supporting the 'strong correlation' claim may be normalization artifacts.","rationale":"I read the paper in good faith: the authors exhaustively simulate all 2^16 input pairs for 20 multiplier designs, and the ED histogram in Figure 1 does show that error varies with inputs, which is genuine evidence for input-dependency. My concern is not that the central claim is false, but that the quantitative evidence for a strong correlation is not yet trustworthy because the NED and PSNR metrics used to visualize the correlation contradict the paper's own equations. The reader's binning concern is valid, but the normalization inconsistency is more load-bearing: if the plotted NED uses local denominators instead of the global Pmax, the apparent dependence on input range may be an artifact rather than a property of the multiplier. This does not require rejecting the paper; the qualitative histogram remains supportive, and the fix is straightforward: recompute the metrics correctly and report the raw per-input error data. I therefore keep the reader's CONDITIONAL verdict, while noting that the concrete test above should be performed before the quantitative claims are relied upon. Secondary concerns include the absence of released code or netlists, which also limits reproducibility, but the metric inconsistency is the decisive issue.","tokens_in":7513,"tokens_out":10608,"duration_ms":108717,"concrete_test":"Recompute every Table 2 entry from raw per-input ED traces using the Section 3.1 formulas: NED = MED/65025 and PSNR = 10*log10(65025/MSE). Two immediate checks: (a) Design1 should read NED=0.00157 and PSNR=5.85 dB instead of 0.0165 and 39.35; (b) recompute the Section 3.3 cluster values with this fixed normalization; if any cluster still exceeds 100.8%, a denominator other than Pmax is being used. If, after correction, the cluster-to-cluster variation and thresholds (NED>=100%, PSNR<25 dB) remain qualitatively similar, the input-dependency claim survives; if the extreme clusters vanish, they were artifacts of local normalization.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim of a strong input-error correlation rests primarily on the NED and PSNR input-dependency surfaces in Sections 3.3 and 3.4. Those surfaces are built from metrics that are internally inconsistent with the definitions in Section 3.1. For Design1 (AMA1/D1), Table 2 lists MED=102, NED=0.0165, MSE=1.69E+04, and PSNR=39.35. Equation (3) with Pmax=(2^8-1)^2=65025 gives NED=102/65025=0.00157, not 0.0165; Equation (7) gives PSNR=10*log10(65025/16900)=5.85 dB, not 39.35. Further, Section 3.3 reports per-cluster NED values above 100% (e.g., 164% for AMA2/D2), which is impossible if each cluster NED is an average of ED/Pmax, since the maximum possible ED is 65535 and Pmax is 65025. This indicates the plotted NED uses a different, unreported denominator (likely a per-cluster maximum or the exact product), and the PSNR values are not derived from the reported MSE. Because the correlation claim is quantified only through these surfaces, a normalization artifact could create the appearance of input-dependency. No correlation coefficient or statistical test is reported, so 'strong correlation' is asserted rather than demonstrated.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript reports an exhaustive accuracy analysis of 20 8-bit approximate array multipliers built from five approximate full-adder types (AMA1-AMA5) and four approximation degrees (D1-D4). It defines ED, MED, NED, RED, MRED, MSE, and PSNR, presents aggregate metrics in Table 2, and visualizes input-dependent behavior through ED histograms and 256-cluster NED/PSNR surfaces. The central claim is that approximation error correlates strongly with the applied input operands, so average metrics can hide large per-input errors. The paper concludes that input-aware quality control could avoid large-error input regimes.","tokens_in":7821,"tokens_out":6949,"duration_ms":65123,"significance":"The exhaustive enumeration over all 2^16 input pairs for 20 designs is a useful empirical resource, and the observation that error varies by input region is plausible and qualitatively visible in the histograms. If the quantitative metrics were corrected and the correlation claim were backed by a statistical measure, the work could guide adaptive quality control in approximate computing. The paper's value is currently limited by internal inconsistencies in the reported metrics and by the absence of any numerical evidence for the claimed 'strong correlation.'","major_comments":[{"comment":"Equation (3) defines NED = MED/Pmax, with Pmax = 65025 for 8-bit multiplication. For Design1 (AMA1/D1), Table 2 lists MED=102 and NED=0.0165, but 102/65025 is approximately 0.00157, not 0.0165. The same discrepancy appears throughout the table (e.g., Design5: 101/65025 is approximately 0.00155, yet NED is listed as 0.0213; Design17: 44/65025 is approximately 0.00068, yet NED is listed as 0.0069). The NED column must be recomputed from Eq. (3) or the definition must be changed explicitly.","section":"Section 3.1 / Table 2, Eq. (3)"},{"comment":"Equation (7) gives PSNR = 10*log10(255^2/MSE). For Design1, with MSE=1.69E+04, this yields 10*log10(65025/16900), which is approximately 5.85 dB, not the reported 39.35 dB. For Design5, MSE=1.44E+04 yields approximately 6.55 dB, not 39.97 dB. The PSNR column is therefore not derived from the reported MSE values; either the MSE column is incorrect or an unreported normalization is being used.","section":"Section 3.1 / Table 2, Eq. (7)"},{"comment":"The reported per-cluster NED values exceed the maximum possible under Eq. (3). Since the largest possible ED between a 16-bit exact product and a 16-bit approximate output is 65535, any cluster-average NED is at most 65535/65025, which is approximately 1.008 (100.8%). The text reports values of 164% for AMA2/D2, 128% for AMA3/D1, and many clusters above 100% for D3/D4 designs. This is impossible if Eq. (3) is applied, implying that the plotted NED uses a different, unreported normalizing denominator. The NED surfaces in Figures 2-6 therefore cannot be interpreted as NED as defined in Section 3.1.","section":"Section 3.3"},{"comment":"The claim of a 'strong correlation' between applied inputs and ED/NED/PSNR is asserted but never quantified. No correlation coefficient, coefficient of determination, or statistical test is reported anywhere; the figures show variation across clusters, but variation alone does not establish correlation, especially when the plotted normalizations are unverified. I request a concrete quantitative measure, such as Pearson or Spearman correlation between input magnitude or input cluster and mean ED/NED, or an R-squared value for cluster means, to support the central claim.","section":"Abstract / Sections 3.2-3.4"},{"comment":"The binning of inputs into blocks of 16 consecutive values is the sole basis for the surfaces in Figures 2-11, yet no sensitivity analysis or formal justification is given. If the error varies sharply within a 16x16 block, the apparent between-cluster structure could be an artifact of the binning rather than a genuine input-dependency. Please report per-input heatmaps or repeat the analysis with at least one alternative cluster size (e.g., 1, 8, 32, 64) to show that the qualitative conclusions are robust.","section":"Section 3.3"}],"minor_comments":[{"comment":"There is a typo in the abstract: 'peak-to-signal-noise ration' should be 'peak-to-signal-noise ratio.'","section":"Abstract"},{"comment":"Equation (7) uses 255^2 while Eq. (3) uses Pmax=(2^n-1)^2; for 8-bit inputs these coincide (65025), but for general n-bit multipliers the relationship should be clarified.","section":"Section 3.1"},{"comment":"The threshold language is confusing: the text first says 'NED >= 20% indicates an unacceptable error' and then uses 'NED <= 100% as our threshold.' Please state the intended acceptability criterion consistently.","section":"Section 3.3"},{"comment":"The figures need axis labels, legends, and colorbars to be interpretable; as presented, the surfaces are referenced but their scales and axes are not fully described.","section":"Figures 2-11"},{"comment":"Reference [1] is incomplete: volume, page numbers, and year are missing.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The numerical inconsistencies in Table 2 and in the per-cluster NED values are severe enough that I would want to see the corrected data and a reproducible calculation before publication. The manuscript is essentially an empirical characterization, so the absence of a statistical test for the central correlation claim is a substantive gap rather than a cosmetic issue. I recommend major revision rather than rejection because the core input-dependency phenomenon is plausible and correctable within the manuscript's scope."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear colleague,\n\nQuick take: the qualitative point is real, but the quantitative evidence for it is broken. The authors exhaustively simulate 20 8-bit approximate array multipliers over all 65536 input pairs and show ED histograms plus per-cluster NED and PSNR surfaces. If taken at face value, those surfaces illustrate that average accuracy metrics like MED and NED hide individual input pairs with much larger error. That warning is useful, and the exhaustive scan is genuine legwork. The reliance on the authors' own earlier designs is fine; self-citation is appropriate here.\n\nThe specific problem is that the reported numbers do not match the paper's own equations. For Design1 D1, Table 2 lists MED=102 and NED=0.0165. Equation (3) with Pmax=65025 gives NED=102/65025=0.00157, not 0.0165. Similarly, Equation (7) with the reported MSE=1.69E4 gives PSNR=5.85 dB, not 39.35. And Section 3.3 reports cluster NED values above 100% — e.g., 164% for AMA2 D2 — which is impossible if NED is ED/Pmax and ED cannot exceed Pmax. So the NED and PSNR surfaces that supposedly demonstrate input-dependency are built on an unreported or different normalization. This is a load-bearing flaw, not a typo.\n\nSecond, the abstract's \"strong correlation\" is asserted but never quantified. There is no correlation coefficient, no conditional expectation, no statistical test. The bar charts show variation, but variation is not correlation. Because ED=|P-P'| by definition, some input-dependency is guaranteed for any nontrivial approximate multiplier; the interesting question is how strong it is and where the bad inputs live. The 16-input binning could also create between-cluster variation if error changes sharply inside a block, and no sensitivity analysis is given. No code or data is provided, so nothing in the quantitative sections is independently reproducible.\n\nWhat the paper does well is present 20 concrete design configurations and explicitly flag input clusters with poor NED or PSNR. That is a potentially useful design-space characterization for approximate-computing researchers, provided the numbers are corrected and the formulas are aligned with the tables.\n\nAs it stands, I would desk-reject this version rather than send it to referees. I'd tell the authors to fix the metric definitions, recompute or clearly re-define NED and PSNR, add a real correlation measure, and release the simulation data. A corrected version could then be worth a workshop or short-paper review; the current one is not.\n\nRecommendation: not referee-ready in its present form.","headline":"The qualitative warning is worth hearing, but the paper's own numbers contradict its metric definitions and the 'strong correlation' claim is unquantified — as submitted it is not referee-ready.","tokens_in":8324,"tokens_out":5455,"would_cite":false,"duration_ms":61358,"reading_group":"no","serious_thinker":"no","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Approximate-multiplier error tracks the specific input operands, not just the design, across all 20 multiplier variants tested.","keywords":["approximate computing","approximate multiplier","error analysis","input dependency","error distance","normalized error distance","peak signal-to-noise ratio","array multiplier"],"falsifier":"Recompute the NED and PSNR maps for Design1 and Design20 with different cluster sizes, such as blocks of 4 or 8, or blocks shifted by one value, and compare per-cluster maxima; if the number and location of clusters that violate the NED ≤ 100% or PSNR ≥ 25 dB thresholds change substantially, the reported correlation is an artifact of the 16-by-16 binning. Alternatively, list the per-input error distance for every one of the 65,536 input pairs of Design1 and check whether the within-block spread in error distance is comparable to the between-block spread.","tokens_in":7316,"feed_emoji":"🔢","tokens_out":8508,"duration_ms":76316,"temperature":0.7,"pith_summary":"An approximate multiplier that looks good on average can still produce unacceptable errors for certain input pairs. This report establishes, by exhaustive simulation of all 65,536 input pairs for each of 20 eight-bit approximate array multipliers, that the error distance, normalized error distance, and peak signal-to-noise ratio of these designs depend strongly on the actual operand values. Across every design in the library, some input clusters violate quality thresholds even when the overall average meets them. The practical point is that average error measures are not enough: an application that can avoid or compensate the high-error input regions could use a cheaper approximate multiplier without sacrificing worst-case quality.","feed_headline":"Inputs, not just design, drive approximate-multiplier error","feed_subtitle":"Averages hide it: in all 20 designs, some input ranges fail quality thresholds even when the mean looks good.","key_machinery":"The central object is a library of 20 designs of 8-bit approximate array multipliers built by combining five types of approximate mirror adders (AMA1–AMA5) with four approximation degrees (D1–D4, which approximate 7, 8, 9, or all 16 bits of the product). The analysis machinery is exhaustive simulation of every possible input pair (65,536 pairs per design), together with a binning step that groups 16 consecutive values of each input into clusters so that NED and PSNR can be averaged per cluster. These per-cluster error maps are what display the input-dependence: each bar in the NED and PSNR plots corresponds to one 16-by-16 block of operand values, and the wide spread across bars is the evidence for the paper's claim.","core_discovery":"The paper's central claim is that approximation error in an 8-bit approximate array multiplier is strongly correlated with the applied inputs, and that this correlation is visible in every one of the 20 designs examined. Concretely, the error distance for a single design can range from zero to hundreds (Design1 spans 0 to 518 around an average of 102), the normalized error distance averaged over clusters of 16 consecutive input values reaches or exceeds 100% for at least some clusters in many designs, and peak signal-to-noise ratio drops below the 25 dB quality threshold for a non-negligible set of input clusters in every design. The report also observes that error magnitude is governed more by the approximation degree, how many result bits are approximated, than by which approximate full-adder type is used, and that average quality metrics conceal these input-dependent spikes. The intended use of the finding is to enable input-aware quality control: by identifying the operand ranges that produce large errors, one can keep those cases out or route them to a more accurate configuration.","pith_inferences":["Beyond the report's binning choice, the 16-by-16 clustering may hide sharper per-input spikes: if errors vary within a block, the reported cluster maps could smooth away the very worst individual input pairs, so recomputing with smaller blocks or per-input error distances would show whether the true error surface is even more input-dependent.","Because the error mechanism is generic to approximate arithmetic, the same input-dependency likely appears in approximate adders, dividers, and tree multipliers, although only array multipliers are tested here.","The observed structure suggests a testable runtime strategy: a small classifier keyed on the high-order operand bits could predict which of the 20 designs will meet a quality target for a given input pair, and the report's own related work on using PSNR as a quality metric points in this direction, but the present report does not build such a controller."],"forward_implications":["Average error metrics such as MED and overall NED underestimate the risk of an approximate multiplier; every design in the library has some input blocks whose NED exceeds 100% or whose PSNR falls below 25 dB.","The error distance approximately doubles each time the approximation degree is increased, so the degree knob is the dominant lever on error magnitude across all five adder types.","A user who can identify high-error input ranges in advance can select a different design or a more accurate mode only for those ranges, keeping the cheap approximate design most of the time.","For D4 designs, where all 16 product bits are approximated, the error is so large that most input clusters are unusable for quality-constrained applications, while D1 designs are usable except for a handful of clusters."],"supporting_citations":[{"why":"Supplies the 20 designed 8-bit approximate array multiplier circuits and the type/degree knobs that this report's analysis is built on.","marker":"[7]"},{"why":"Provides the five approximate mirror adders (AMA1–AMA5) that form the type knob of the multiplier library.","marker":"[8]"},{"why":"Definition source for the error metrics (ED, MED, NED, RED, MRED, MSE, PSNR) used throughout the analysis.","marker":"[10]"},{"why":"Gives the earlier proposal to use PSNR as a quality-of-result metric for controlling approximate computing, which motivates the PSNR analysis here.","marker":"[11]"},{"why":"Supplies the PSNR ≥ 25 dB threshold used to classify output quality as acceptable.","marker":"[12]"}],"fun_headline_variants":["Approximate multiplier error spikes follow input ranges","Average error masks input-dependent spikes in 20 multipliers","Approximation degree sets error magnitude more than adder type","All 20 approximate multiplier designs show input-dependent error"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The analysis assumes that grouping every 16 consecutive input values into a cluster faithfully represents how error varies across inputs; if errors change sharply inside a 16-by-16 block, the strong input-dependence shown in the NED and PSNR plots could be a binning artifact rather than a genuine property of the multiplier.","fun_headline_variants_meta":{"raw":{"variants":["Approximate multiplier error spikes follow input ranges","Average error masks input-dependent spikes in 20 multipliers","Approximation degree sets error magnitude more than adder type","All 20 approximate multiplier designs show input-dependent error"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000915,"raw_usage":{"total_tokens":3918,"prompt_tokens":925,"completion_tokens":2993,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":541,"completion_tokens_details":{"reasoning_tokens":2941}},"tokens_in":541,"tokens_out":2993,"duration_ms":23811,"temperature":1.0,"reasoning_tokens":2941,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T15:16:09.265374+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Recompute the NED and PSNR maps for Design1 and Design20 with different cluster sizes, such as blocks of 4 or 8, or blocks shifted by one value, and compare per-cluster maxima; if the number and location of clusters that violate the NED ≤ 100% or PSNR ≥ 25 dB thresholds change substantially, the reported correlation is an artifact of the 16-by-16 binning. Alternatively, list the per-input error distance for every one of the 65,536 input pairs of Design1 and check whether the within-block spread in error distance is comparable to the between-block spread.","supporting_citations":[{"cited_title":"Comparative study of approximate multi- pliers,","cited_arxiv_id":null,"evidence_quote":"Supplies the 20 designed 8-bit approximate array multiplier circuits and the type/degree knobs that this report's analysis is built on."},{"cited_title":"Low-power digital signal processing using approximate adders,","cited_arxiv_id":null,"evidence_quote":"Provides the five approximate mirror adders (AMA1–AMA5) that form the type knob of the multiplier library."},{"cited_title":"Inexact designs for approx- imate low power addition by cell replacement,","cited_arxiv_id":null,"evidence_quote":"Definition source for the error metrics (ED, MED, NED, RED, MRED, MSE, PSNR) used throughout the analysis."},{"cited_title":"Using machine learning for quality con- figurable approximate computing,","cited_arxiv_id":null,"evidence_quote":"Gives the earlier proposal to use PSNR as a quality-of-result metric for controlling approximate computing, which motivates the PSNR analysis here."},{"cited_title":"Quality of service for video streaming over multi-hop wireless networks: Admission control approach based on analytical capacity estimation,","cited_arxiv_id":null,"evidence_quote":"Supplies the PSNR ≥ 25 dB threshold used to classify output quality as acceptable."}],"review_version":1}