{"id":"51031486-dfcc-42e2-9cf5-8449f5fbb7e8","arxiv_id":"2501.07564","paper_version":2,"verdict":"REJECT","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":1,"one_line_summary":"E2ESlack predicts pre-routing slack and TNS/WNS via a GNN arrival-time model plus a heuristic required-time calculator, yet its headline results are computed with ground-truth arrival times rather than the model's predictions.","lead":"This paper presents E2ESlack, a framework that predicts digital circuit timing slacks before routing by combining a graph neural network for arrival times with a fast required-time estimator. It reports TNS/WNS close to post-routing static timing analysis and up to 23x faster runtime, but the reported accuracy assumes a perfect arrival-time predictor.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The end-to-end TNS/WNS claim is untested: Tables 3 and 4 use post-routing STA AT labels, with the paper explicitly assuming the AT prediction model is ideal, so the reported agreement with post-routing STA does not validate the actual E2ESlack pipeline.","rationale":"The reader identified the weakest assumption as the use of ground-truth AT labels in the TNS/WNS evaluation, and this is indeed the single most load-bearing issue. The abstract and contributions claim an end-to-end framework that predicts slack and TNS/WNS at the pre-routing stage with runtime savings, but the accuracy experiments in Section 5.2 explicitly substitute post-routing STA AT labels for the framework's predicted AT. That substitution removes the main source of error in the end-to-end pipeline from the accuracy measurement. Because slack is defined as RAT minus AT, and both WNS and TNS depend on the sign and magnitude of endpoint slacks, AT prediction errors can change which paths are critical and by how much. The paper even acknowledges TNS is sensitive to accumulated errors, but only demonstrates this under ideal AT. The runtime experiments do include AT prediction, so the reported speedup is honest, but the accuracy claim is not end-to-end. The paper deserves credit for clearly disclosing the assumption in Section 3.3, and the RAT estimation algorithm itself may be a useful contribution evaluated fairly under ideal AT. However, as published, the headline claim that E2ESlack predicts TNS/WNS comparably to post-routing STA is unsupported. My assessment is a rejection with the possibility of acceptance after a revised version reports end-to-end TNS/WNS results computed from predicted AT and, ideally, releases code and data so the community can reproduce those results.","tokens_in":11251,"tokens_out":2272,"duration_ms":25455,"concrete_test":"Run the complete E2ESlack inference pipeline on the 11 test circuits: TimingParser graph construction, TimingPredict AT prediction, Algorithm 1 RAT estimation, and TNS/WNS computation using the model-predicted AT at endpoints (not the post-routing STA AT labels). Compare these end-to-end TNS/WNS values against the post-routing STA ground truth used in Table 4. Report endpoint slack MAE, TNS/WNS absolute differences, and the number of endpoints whose slack sign flips relative to ground truth. If the resulting TNS/WNS errors remain as small as the ideal-AT errors in Table 4, the end-to-end claim survives; if not, the paper must be revised to claim only RAT estimation accuracy, not end-to-end pre-routing slack prediction.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim is that E2ESlack 'achieves TNS/WNS values comparable to post-routing STA results while saving up to 23x runtime.' Slack is defined as RAT - AT_pred, so the end-to-end claim requires the framework's own AT predictions to be accurate enough that endpoint slacks, TNS, and WNS remain close to post-routing STA. This condition is never tested. In Section 3.3, the paper states verbatim: 'when presenting RAT/TNS/WNS results of our algorithm in Table 4 and Table 3 we use AT labels (from post-routing STA) instead of predicted AT from TimingPredict. That is, we assume that the AT prediction model is ideal.' Consequently, Table 4's 'Ours' TNS/WNS columns are generated with ground-truth AT labels; they validate the RAT estimation algorithm plus an assumption of perfect AT, not the end-to-end framework. This is load-bearing because AT prediction error on a few critical endpoints can change the sign of small-margin slacks, and both WNS and TNS are highly nonlinear functions of endpoint slack. The reported test-set AT R2 averages around 0.89, but no analysis is given of AT error at endpoint pins that determine negative slack, so the magnitude of the induced TNS/WNS error is unknown. The runtime comparison, by contrast, does include the full TO parsing, AT prediction, and RAT estimation pipeline (Section 5.3). Mixing an idealized-AT accuracy measurement with a full-pipeline runtime measurement supports neither the 'end-to-end' nor the '23x with comparable accuracy' statement as written. The paper honestly discloses this limitation, but the disclosure identifies the exact gap that invalidates the headline contribution as stated. A revised version could fix this by reporting TNS/WNS computed from predicted AT, or by explicitly reframing the contribution as RAT estimation under ideal AT.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes E2ESlack, an end-to-end framework for pre-routing slack prediction. It consists of a distributed TimingParser that converts LIB/SDF/DEF files into DGL graphs, a GNN-based arrival time (AT) prediction model based on TimingPredict, and a deterministic RAT estimation algorithm that computes endpoint RATs from the clock period, SDC constraints, clock-path delays, and clock reconvergence pessimism (CRPR). From predicted AT and estimated RAT, the framework computes endpoint slacks and TNS/WNS. The authors evaluate AT prediction R2, RAT MAE against modified TimingPredict and pre-routing STA, TNS/WNS against post-routing STA, and runtime speedup. A central caveat stated in Section 3.3 is that all RAT/TNS/WNS results use ground-truth post-routing AT labels rather than the framework's own AT predictions.","tokens_in":11602,"tokens_out":5364,"duration_ms":52032,"significance":"If the claimed performance held for the actual end-to-end pipeline, this would be a practically useful shift-left tool: pre-routing TNS/WNS indicators at large runtime savings would enable earlier timing closure iterations. The modular parser and the CRPR-aware RAT estimation idea are useful contributions, and the runtime comparison includes full pipeline steps that prior work neglected. However, the main claim is currently not supported by the experiments, because accuracy is measured under an ideal-AT assumption while runtime is measured on the full pipeline; moreover, even under that assumption the TNS/WNS agreement is poor on several circuits.","major_comments":[{"comment":"The central claim that E2ESlack achieves TNS/WNS values comparable to post-routing STA is not tested with the framework's own AT predictions. The paper states verbatim that RAT/TNS/WNS results use AT labels from post-routing STA and assume the AT prediction model is ideal. Since slack is defined as RAT minus AT_pred, the end-to-end claim requires the predicted AT to be accurate at critical endpoints; no experiment reports TNS/WNS with predicted AT. Thus Tables 3 and 4 validate the RAT formula conditional on perfect AT, not the proposed end-to-end framework.","section":"Section 3.3, Tables 3 and 4"},{"comment":"Even under the ideal-AT assumption, the TNS/WNS results do not support the claim of comparability to post-routing STA. For des, TNS is -1945.19 versus -18.41 and WNS is -5.62 versus -1.18; for y_huff, TNS is -5.17 versus 0.00; for xtea, TNS is -15.13 versus -1.14. These are order-of-magnitude TNS discrepancies and WNS errors of several nanoseconds. The low average RAT MAE in Table 3 does not preclude large errors on the few endpoints that determine TNS/WNS, because MAE is averaged over all endpoints. The paper should report per-endpoint slack error distributions and critical-endpoint error statistics.","section":"Table 4"},{"comment":"The RAT comparison is not apples-to-apples: \"Ours\" uses ground-truth post-routing AT labels, while the modified TimingPredict baseline must predict RAT from circuit features and the pre-routing STA baseline uses no routing parasitics. The claimed 57.4% average MAE reduction is therefore a comparison of an oracle-assisted formula against prediction/tool baselines. A fair comparison would feed the same information, such as predicted AT from TimingPredict, into all RAT estimators, or would at minimum report RAT MAE conditional on predicted AT.","section":"Section 5.1, Table 3"},{"comment":"The runtime experiment measures the full pipeline including parser, AT prediction, and RAT estimation, but the accuracy experiment assumes an ideal AT model. Consequently, no single configuration is shown to simultaneously achieve the reported accuracy and the 23x speedup. The \"comparable accuracy with 23x runtime\" claim in the abstract and introduction mixes two different evaluation setups.","section":"Section 5.3 versus Section 3.3"}],"minor_comments":[{"comment":"The paper only considers late/setup slack, but the abstract and title use the unqualified term \"slack prediction\"; please state the corner limitation in the abstract or contributions.","section":"Section 2.1"},{"comment":"The formulas for RAT_PO and RAT_S in the text are poorly typeset and the treatment of the setup time term Tsu is confusing: it is mentioned in the text but then ignored in the estimation. Please rewrite these definitions in a clear equation block and justify the approximation quantitatively.","section":"Section 3.3"},{"comment":"The speedup column shows \"0x\" for several small circuits; please clarify whether this means less than 1x speedup, rounding, or a missing value.","section":"Table 3"},{"comment":"No URL or repository link is provided for the new OpenLane benchmark circuits or the exact train/test split; without this information the results are hard to reproduce.","section":"Section 4.2"},{"comment":"The ACM reference format line \"InProceedings of (Arxiv)\" is malformed and should be corrected.","section":"References"},{"comment":"The phrase \"modified TimingPredict\" is used for a model that is retrained to predict RAT in addition to AT; please define this modification precisely, including the loss function and training setup, before using it as a baseline.","section":"Section 5.1"}],"recommendation":"reject","confidential_remarks":"The paper has a useful modular framework and a sensible RAT estimation idea, and it is honest about the ideal-AT assumption. However, the headline claims are unsupported: the accuracy evaluation is oracle-assisted while the runtime evaluation is full-pipeline, and even the oracle-assisted TNS/WNS results show large discrepancies on several circuits. This is not a presentation issue; the experiments as designed cannot validate the end-to-end claim. If the authors can provide a full-pipeline evaluation with predicted AT and show acceptable critical-endpoint error distributions, a resubmission could be considered."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: The RAT estimation module is a real, useful contribution, and the paper is honest about its own limits. But the abstract's end-to-end claim—E2ESlack achieves TNS/WNS comparable to post-routing STA—is not supported by the experiments as run. Section 3.3 says plainly that the TNS/WNS and RAT results use ground-truth AT labels, i.e., assume the AT predictor is ideal. So Table 4 validates the RAT algorithm, not the full pipeline.\n\nWhat's new: the first pre-routing RAT estimation that handles non-PO endpoints via clock-path delay estimates and CRP correction. The comparison to a modified TimingPredict and pre-routing STA shows genuine improvement in RAT MAE when AT is assumed perfect. The TimingParser is also more complete than prior work—it parses DEF/SDF/LIB and is distributed. The runtime analysis includes parsing, AT inference, and RAT estimation, which is fairer than TimingPredict's inference-only timing.\n\nSoft spots: the headline overreach. AT R2 of 0.89 on test does not guarantee small slack errors; TNS/WNS are dominated by a few critical endpoints, and a small AT error can flip a negative slack to positive or vice versa. The paper even acknowledges TNS prediction is 'less accurate' but the abstract still claims comparability. Also, the RAT formula itself is standard STA (clock period minus uncertainty plus CRP), so the novelty is in application and CRP approximation rather than a new timing theory. The comparison to 'pre-routing STA' is a bit odd—they run STA without SPEF, which is known to be inaccurate, and their method uses ground-truth AT, so it's not a fully fair apples-to-apples comparison.\n\nWho this is for: EDA researchers working on pre-routing timing prediction. The RAT estimation part is worth building on. The paper deserves peer review: a serious referee would ask for end-to-end results with predicted AT, and the authors could likely provide them. The disclosed limitation is a fixable gap, not a fatal flaw.\n\nRecommendation: send it to review, but flag the evaluation gap clearly and require either end-to-end TNS/WNS or a revised claim that scopes the contribution to RAT estimation under ideal AT.","headline":"A genuinely useful RAT estimation module under ideal AT, but the end-to-end slack claim is untested and the abstract overstates it.","tokens_in":12220,"tokens_out":3077,"would_cite":false,"duration_ms":26634,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"E2ESlack claims to be the first framework that predicts path-level timing slacks before routing, from placement-stage circuit data, with TNS/WNS close to post-routing static timing analysis and up to 23x runtime savings.","keywords":["pre-routing slack prediction","required arrival time estimation","graph neural network","static timing analysis","TNS/WNS","timing closure","placement-stage timing","EDA"],"falsifier":"Re-run the TNS/WNS evaluation on the same test circuits using the framework's predicted arrival-time values instead of the post-routing STA labels; if TNS errors become comparable in magnitude to the true negative slacks, or if the set of predicted critical paths diverges from ground truth, the end-to-end pre-routing slack claim would be refuted.","tokens_in":1643,"feed_emoji":"⏱️","tokens_out":1877,"duration_ms":75942,"temperature":0.7,"pith_summary":"E2ESlack tries to establish that circuit timing quality—the total and worst negative slack (TNS/WNS) metrics that decide whether a chip meets its clock constraints—can be obtained before routing, using placement-stage circuit files and design constraints. The method combines a parser that turns standard circuit formats into pin-level graphs, a graph-neural-network arrival-time predictor, and a new fast algorithm for estimating required arrival time (RAT). With RAT available, setup slack is computed as RAT minus arrival time, and TNS/WNS follow by aggregating negative-slack endpoints. The authors report that their RAT estimates are more accurate than pre-routing static timing analysis and an adapted prior AT-model baseline, and that the overall pipeline produces TNS/WNS close to post-routing STA while saving up to 23x runtime. If this holds, designers could catch timing violations and iterate on placement long before the expensive routing step.","feed_headline":"Pre-routing slack prediction cuts timing checks by 23x","feed_subtitle":"A placement-stage estimator yields TNS/WNS close to post-routing static timing analysis, so designers catch violations before routing.","key_machinery":"The load-bearing mechanism is the RAT estimation module: an algorithm that converts SDC timing constraints and predicted arrival times into a required arrival time for every timing endpoint. For a primary output, endpoint RAT is set to the clock period minus the specified output delay and clock uncertainty; for a register endpoint, the early clock-path delay is estimated from the predicted arrival time at the flip-flop's clock pin, and a clock-reconvergence pessimism correction is added—the difference between the maximum and minimum delay of the common launch-and-capture clock path—but only after a cheap negative-slack path traversal. This yields slack = RAT − AT at each endpoint and therefore TNS/WNS, closing the gap that earlier arrival-time-only predictors left open. The framework also includes a graph parser that turns standard placement, netlist, and library files into pin-node graphs with timing lookup-table features, with variable-sized lookup tables interpolated to a fixed size.","core_discovery":"The central claim is that path-level slack prediction at the pre-routing stage is achievable end-to-end: given a placed but unrouted circuit, the framework constructs a heterogeneous pin-level graph, predicts arrival times with a timing-engine-inspired graph neural network, and estimates required arrival times at every endpoint from the clock period, SDC constraints, and predicted arrivals on clock pins. It then computes setup slack as RAT minus arrival time, identifies negative-slack critical paths, adds a clock-reconvergence-pessimism correction for those paths, and aggregates the results into TNS and WNS. The authors state that this is the first work addressing pre-routing RAT estimation and path-level slack prediction, and they report a test-set RAT mean absolute error of 0.6454 versus 1.5165 for pre-routing STA and 8.1608 for the adapted prior arrival-time model. They also report TNS/WNS values close to post-routing STA and up to 23x runtime savings. As the paper explicitly notes, the RAT and TNS/WNS numbers are computed using ground-truth post-routing arrival-time labels, so those results assume the arrival-time prediction model is ideal.","pith_inferences":["Editorial extension: if the ideal-arrival-time assumption holds, the most natural next step is to use E2ESlack inside a placement-optimization loop as a cheap timing surrogate, since the reported 23x saving makes it fast enough to evaluate many placement candidates.","Editorial extension: the paper leaves untested whether the RAT estimates remain accurate under early/hold corner analysis, multi-clock or gated-clock designs, or process corners beyond the late-rise setup corner; these are natural stress tests for the algorithm.","Editorial extension: an end-to-end validation using the framework's own predicted arrival times, rather than post-routing STA labels, would be needed to know whether the reported TNS/WNS closeness persists; the authors' own assumption of an ideal AT predictor marks the boundary between the pipeline as described and the pipeline as evaluated."],"forward_implications":["Designers can obtain early TNS/WNS estimates during placement, before routing, and use them to decide whether a design should be sent to the expensive routing step.","Any future arrival-time prediction model can be plugged into the framework; the RAT module turns predicted arrival times into path-level slack, so improvements in arrival-time accuracy directly improve RAT and TNS/WNS estimates.","RAT estimation is faster and more accurate than pre-routing STA on the tested circuits, meaning a pre-routing STA step can be replaced by the graph-based estimate.","The framework reduces the time needed to obtain TNS/WNS metrics by up to 23x compared with routing plus post-routing STA, enabling faster design iterations and earlier timing-closure decisions.","Because the paper reports that TNS prediction accumulates error across many paths while WNS tracks more closely, TNS estimates are best treated as a screening signal rather than exact signoff numbers."],"supporting_citations":[{"why":"Establishes the earlier ML-based pre-routing timing prediction line that this work extends to full slack.","marker":"[1]"},{"why":"Supplies the post-routing STA ground truth and the pre-routing STA baseline used for comparison.","marker":"[4]"},{"why":"Provides the OCV and clock-reconvergence-pessimism timing semantics that motivate the CRP correction in the RAT estimation.","marker":"[5]"},{"why":"Supplies the modified STA dump that includes arrival time, RAT, and slew labels in SDF for training the arrival-time model.","marker":"[8]"},{"why":"Supplies the arrival-time prediction model and graph representation reused as-is, along with the original benchmark circuits and train-test split.","marker":"[9]"},{"why":"Supplies the additional open benchmark circuits used to build the new dataset extension.","marker":"[19]"},{"why":"Represents the prior arrival-time-only pre-routing prediction work that lacks an endpoint RAT mechanism, defining the gap this paper addresses.","marker":"[22]"}],"fun_headline_variants":["End-to-end pre-routing slack predictor delivers 23x speedup","First graph framework predicts pre-routing path slacks, 23x runtime gain","Pre-routing slack prediction matches post-routing STA, 23x faster","E2ESlack: end-to-end graph framework for pre-routing slack, 23x speedup"],"cache_read_input_tokens":14208,"weakest_assumption_plain":"The paper's headline RAT and TNS/WNS numbers are calculated using true post-routing arrival times rather than the framework's own predicted arrivals, so the end-to-end accuracy claim rests on the unverified assumption that arrival-time prediction errors are small enough not to change the slack conclusions.","fun_headline_variants_meta":{"raw":{"variants":["End-to-end pre-routing slack predictor delivers 23x speedup","First graph framework predicts pre-routing path slacks, 23x runtime gain","Pre-routing slack prediction matches post-routing STA, 23x faster","E2ESlack: end-to-end graph framework for pre-routing slack, 23x speedup"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001184,"raw_usage":{"total_tokens":4924,"prompt_tokens":1015,"completion_tokens":3909,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":631,"completion_tokens_details":{"reasoning_tokens":3820}},"tokens_in":631,"tokens_out":3909,"duration_ms":25785,"temperature":1.0,"reasoning_tokens":3820,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T20:38:17.741602+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the TNS/WNS evaluation on the same test circuits using the framework's predicted arrival-time values instead of the post-routing STA labels; if TNS errors become comparable in magnitude to the true negative slacks, or if the set of predicted critical paths diverges from ground truth, the end-to-end pre-routing slack claim would be refuted.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the post-routing STA ground truth and the pre-routing STA baseline used for comparison."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the OCV and clock-reconvergence-pessimism timing semantics that motivate the CRP correction in the RAT estimation."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the modified STA dump that includes arrival time, RAT, and slew labels in SDF for training the arrival-time model."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the arrival-time prediction model and graph representation reused as-is, along with the original benchmark circuits and train-test split."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the additional open benchmark circuits used to build the new dataset extension."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Represents the prior arrival-time-only pre-routing prediction work that lacks an endpoint RAT mechanism, defining the gap this paper addresses."}],"review_version":1}