{"id":"67a799e4-2379-4390-a614-fecbdd425a90","arxiv_id":"2506.21371","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"Fine-grained assignment of approximate multipliers across layers, filters, and kernels of ResNet-8 on CIFAR-10 yields up to 54% energy savings with 4% accuracy loss.","lead":"This paper shows a way to save up to 54% of the energy used by a small image-recognition chip by making some of its multiplications slightly less exact, while keeping accuracy loss to about 4%. It could help battery-powered devices run neural networks more efficiently.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The headline energy savings rest on an energy model (#mult x avg multiplier energy) that omits memory, control, and the hardware overhead of heterogeneous per-filter/per-kernel multiplier placement; without a full-accelerator measurement the central energy claim is unverified.","rationale":"Reading the manuscript in good faith, the accuracy side of the claim is internally supported: Table I reports accuracy losses of 17-21% relative to the full precise model, which translate to 0-4% relative to the quantized baseline, so the 'up to 4% accuracy loss' statement is consistent. The comparison against EvoApprox8b configurations is a reasonable external baseline, and the exploration of layer/filter/kernel granularity is a genuine extension of ALWANN. The main weakness is exactly where the reader placed it: the energy model in Section III-F is too coarse to validate system-level energy claims. I would sharpen the reader's formulation: the missing terms are not only memory and control, but also the overhead of physically realizing heterogeneous multiplier types at filter and kernel granularity, which is the paper's central technical novelty. If a real accelerator must pay area or routing costs to place many different approximate multipliers, the per-multiplier energy savings will not fully translate. The correct response is not to reject the paper but to condition acceptance on a full-accelerator energy measurement or, at minimum, an explicit statement that the energy results are multiplier-level upper bounds. The paper should also correct the '2x gains' phrase in the abstract and conclusion, since Table I supports roughly 1.4x at best. These issues do not invalidate the methodology or the accuracy results, so the conditional verdict remains appropriate.","tokens_in":6637,"tokens_out":5099,"duration_ms":57489,"concrete_test":"Implement one reported Pareto-optimal configuration, e.g., KLAM-chan 2 0 2 from Table I, as a complete RTL accelerator using the same TSMC 45nm standard-cell flow: generate a heterogeneous PE array with the assigned ROUP multipliers, include weight/activation buffers, data routing, and controller logic, and run Synopsys PrimeTime at the same corner used for the individual multipliers. Partition the measured total energy into multiplier, memory, and interconnect components. If the system-level energy reduction versus the exact-multiplier baseline is within 10 percentage points of the Section III-F estimate, the energy claim stands; if the measured gain is materially lower (e.g., less than half the estimated reduction), the headline savings are not realizable in a practical accelerator.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central contribution is that fine-grained assignment of ROUP approximate multipliers yields 46-54% energy gains with at most 4% accuracy loss compared to the quantized baseline. The energy estimate that supports this claim is defined in Section III-F as the sum over layers of #mult x avg multiplier energy, with no contribution from memory access, data movement, or control logic. In a real DNN accelerator, multiplier datapath energy is only one component; weight/activation storage and data movement typically consume a substantial share of total energy, especially in 45nm ASIC implementations. If multipliers are not the dominant energy consumer, the reported reductions will be diluted at the system level. The concern is sharpened by the paper's own novelty: unlike ALWANN, which uses one multiplier type per layer, MAx-DNN assigns different ROUP multipliers to individual filters or kernels. A physical accelerator implementing this heterogeneity must instantiate multiple multiplier designs in the processing array, or add selection/routing logic, and must handle scheduling and control for the mixed multiply operations. The #mult x avg model implicitly assumes this heterogeneity is free. None of these costs are quantified, so the energy savings are an optimistic upper bound rather than a measured system-level result. The '2x energy gains versus EvoApprox8b' statement is also not supported by Table I, where the maximum gains are 54% versus 39%, i.e., roughly 1.38x, but this overstatement is secondary to the energy-model gap.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes MAx-DNN, an extension of the ALWANN framework that assigns approximate ROUP multipliers to DNN multiplications at three granularities: layer-level (LLAM), filter-level (FLAM), and kernel-level (KLAM, with channel/row/column variants), plus a multiplication-skip variant (KLMS). Evaluation on quantized ResNet-8 and CIFAR-10 reports up to 54% energy gain with at most 4% accuracy loss relative to the quantized baseline, and claims 2x energy efficiency compared to ALWANN with EvoApprox8b multipliers. The central contribution is the fine-grained, non-uniform distribution of approximate multipliers and the accompanying design-space exploration.","tokens_in":6934,"tokens_out":5431,"duration_ms":55665,"significance":"If the energy numbers are taken at face value, fine-grained multiplier approximation is a practical way to trade a few percent accuracy for substantial energy savings in DNN inference accelerators. The use of ROUP multipliers with a dense error-energy trade-off is sensible, and the controlled comparison in Fig. 4 between LLAM, FLAM, and KLAM on the same multiplier library is a useful result for the approximate-computing community. However, the headline system-level energy savings rest on a simplified energy model, and the advertised 2x advantage over the EvoApprox8b baseline is not directly supported by the data. The paper is a worthwhile proposal but needs additional validation before its claims can be considered established.","major_comments":[{"comment":"The energy estimate in Section III-F is defined as #mult × avg multiplier energy per layer, with no contribution from memory access, data movement, control logic, or the overhead of instantiating heterogeneous multiplier types within a processing array. Since the paper's central claim is energy reduction, this model is load-bearing; if multiplier datapath energy is not dominant in a real accelerator, the reported 46–54% gains will be diluted. The paper should either report a more complete energy model including memory and control, or present measured results from a synthesized accelerator implementing the heterogeneous assignments, or explicitly caveat the headline as an upper bound.","section":"Section III-F"},{"comment":"The claim of '2× energy gains versus the state-of-the-art DNN approximations' is not supported by Table I. The maximum proposed energy gain is 54%, while the best EvoApprox8b configuration (Evo mul8u ZFB) has a 39% gain, a ratio of roughly 1.4×. Moreover, the comparison is confounded: the proposed configurations use ROUP multipliers with fine-grained assignment, while the baseline uses EvoApprox8b multipliers with layer-wise assignment, so any improvement could stem from the multiplier library rather than from the fine-grained level. The controlled comparison in Fig. 4 (all ROUP) supports the fine-grained benefit, but the abstract and conclusion should not attribute the 2× claim to the approximation level without a same-library comparison.","section":"Table I, Section IV"},{"comment":"KLMS (kernel-level multiplication skip) is defined as a proposed approach, but it is never evaluated separately. Fig. 4 groups 'KLAM+KLMS' and Table I lists only KLAM configurations, so the contribution of the skip mechanism is unquantified. The reader cannot determine whether KLMS improves the Pareto front or whether it is an independent design choice.","section":"Section III-D, Fig. 4"},{"comment":"The configuration labels in Table I (e.g., 'FLAM-3clas. 2 1 1', 'KLAM-chan. 1 0 1') are not explained, and the ROUP L/M/H multipliers used in the sensitivity study are not mapped to their (P, r) parameters. Without this information, the reported Pareto-optimal configurations cannot be reproduced or compared across papers, which undercuts the empirical contribution.","section":"Table I, Section IV"}],"minor_comments":[{"comment":"The word 'defacto' should be 'de facto', and 'state-of-the-art' should be hyphenated consistently.","section":"Abstract"},{"comment":"The baseline label is 'm=0' in part (a) but 'Layer=0' in the caption; use a single convention throughout.","section":"Figure 3"},{"comment":"The caption states that 'Accuracy Loss' is compared to the full accurate model, but the text refers to 4% loss relative to the quantized baseline. Please define the reference point in the caption and in the text to avoid confusion.","section":"Table I"},{"comment":"The energy model is described only in prose; include an explicit equation and state the assumptions about input statistics and switching activity used for the power measurements.","section":"Section III-F"},{"comment":"The sentence 'the average loss of the EvoApprox8b configurations is ∼23%' should specify that this is relative to the full accurate model, otherwise it appears inconsistent with the 4% claim in the abstract.","section":"Section IV"},{"comment":"The framework name ALWANN is typeset as 'ALW ANN' with a space in several places (e.g., Section II and Fig. 1); use 'ALWANN' consistently.","section":"Throughout"}],"recommendation":"major_revision","confidential_remarks":"The manuscript appears to be the authors' version of a LASCAS 2022 conference paper now posted on arXiv. The core idea is relevant, but the journal-level bar requires a more complete energy model, a controlled comparison against the EvoApprox8b baseline, and an ablation of KLMS. The novelty disclosure is adequate: ROUP and ALWANN are properly cited, and the extension is clearly positioned."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Let me give you the short version: this is a small but legitimate extension of ALWANN that demonstrates a real point—fine-grained assignment of approximate multipliers (per filter or kernel) gives a better energy-accuracy Pareto front than layer-level assignment. The experiments are internally consistent and the comparison to the external EvoApprox8b library is meaningful. But the headline energy numbers are computed with a simplified model that omits memory, control, and the overhead of heterogeneous multiplier placement, and the paper overstates one result: the '2x energy gains' vs ALWANN is actually about 1.3x from Table I.\n\nWhat's new: the FLAM and KLAM distribution schemes, and the systematic Pareto comparison showing that finer granularity beats coarser granularity. That is a useful data point for anyone designing approximate DNN accelerators. The paper also does a reasonable job of describing the toolflow, and the ROUP multiplier library is properly characterized in prior work. No retraining is needed, which is a practical plus.\n\nThe soft spots: The energy model (Section III-F) is just #multiplications × average multiplier energy per layer. That ignores weight/activation storage, data movement, and the extra logic needed to support different multipliers in the same processing array. So the 'up to 54% energy gain vs quantized baseline' is really a multiplier-level saving, not a system-level savings. For the main comparative claim—filter/kernel beats layer—this is less of a problem because the same model is used for all approaches. More annoying: the configuration labels in Table I (e.g., 'FLAM-3clas. 2 1 1') are never explained, so a reader can't reproduce the configurations. There are no error bars on accuracy or energy. KLMS is introduced as a fourth approach but never evaluated separately; the Pareto plot merges KLAM+KLMS. And the '2x energy gains' phrase in the abstract and text is not supported: Table I shows max 54% vs 39% energy gain, about 1.3x. That should be corrected.\n\nBottom line: this is a credible conference-level result, modest in scope. It's worth a serious referee as a workshop/short-paper quality contribution, but the energy claims need to be positioned as component-level, and the performance statements need to match the table. If you're working on approximate DNN accelerators, it's a worthwhile read for the granularity finding.","headline":"Fine-grained multiplier assignment shows real Pareto gains, but the energy model is component-level and one claim is overstated.","tokens_in":7503,"tokens_out":3222,"would_cite":false,"duration_ms":33025,"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":"Distributing approximate multipliers at filter and kernel granularity cuts DNN inference energy by up to 54% while keeping accuracy loss to 4% or less.","keywords":["Approximate computing","Inexact multipliers","Energy-efficient hardware accelerators","Deep neural networks","ResNet-8","CIFAR-10","Filter-level approximation","Kernel-level approximation"],"falsifier":"Build one of the reported Pareto-optimal approximate accelerators with a realistic memory hierarchy and measure total inference energy against the exact quantized accelerator; if multiplier energy is a small share of total system energy, the claimed 54% gain will not reproduce.","tokens_in":6436,"feed_emoji":"⚡","tokens_out":14214,"duration_ms":139775,"temperature":0.7,"pith_summary":"MAx-DNN asks where, exactly, a convolutional network can absorb arithmetic approximation: which layers, which filters, and which kernels should use which inexact multiplier. On quantized ResNet-8 trained on CIFAR-10, the paper's layer-, filter-, and kernel-level allocation schemes report up to 54% multiplication-energy savings with no more than 4% accuracy loss, and filter- and kernel-level placements beat layer-uniform approximation by roughly two times in energy with better accuracy. A layer-sensitivity study supports the placement logic: the first convolutional layers are fragile, while later layers absorb much stronger approximations. The intended consequence is a no-retraining design path from frozen quantized models to lower-power approximate hardware accelerators.","feed_headline":"Inexact multipliers per filter or kernel cut DNN energy 54%","feed_subtitle":"At only 4% accuracy loss compared with the quantized model, fine-grained placement beats uniform layer approximation.","key_machinery":"The load-bearing mechanism is the ROUP approximate multiplier family, whose two tunable operations, partial-product perforation and asymmetric rounding, create a dense error-energy trade-off. For N-bit inputs the approximate product is $$\\mathrm{ROUP}(A,B)=\\sum_{j=P}^{N/2-1}\\tilde{P}_j4^j,$$ where P is the number of least-significant partial products omitted and $\\tilde{P}_j$ is a partial product whose operand is rounded to a different bit-width per column. MAx-DNN turns this library into a placement problem: an extended convolution operator lets each layer, filter group, or kernel carry its own multiplier choice, and a Pareto-optimal search selects combinations on the accuracy-energy plane. The governing energy estimate is per-layer multiplication count times average multiplier energy accumulated over the network, which is why the multiplier library dominates the comparison.","core_discovery":"The paper's central claim is that arithmetic approximation for DNN accelerators should be allocated heterogeneously at the finest practical granularity. MAx-DNN assigns a different ROUP approximate multiplier to each convolutional layer, to groups of filters, or to individual kernels, with separate channel, row, and column flavors, and in the KLMS variant skips multiplications whose kernel weights lie inside a chosen band around the layer mean. Compared with the quantized baseline, the reported best configurations save up to 54% of multiplication energy at up to 4% accuracy loss; compared with configurations that apply the same approximate multipliers uniformly per layer, the filter- and kernel-level designs report about two times lower energy and higher accuracy, and they form almost the entire accuracy-energy Pareto front. The layer-wise sensitivity experiment completes the story: the first convolutional layers are markedly less error-resilient than the final ones.","pith_inferences":["Running the same placement search jointly with quantization-aware training or pruning of the multiplier choices could push the accuracy-energy frontier further; the paper fixes the quantized model and only varies multipliers.","The fragility of the first layers suggests error amplification through shared features and residual connections, so the sensitivity profile may shift on non-residual or deeper networks, a prediction that a repeat of the layer sweep could test.","On process nodes where data movement dominates, the likely next question is whether approximation should be restricted to compute-bound layers while memory traffic stays exact; the paper's multiplier-only energy estimate leaves that scenario open."],"forward_implications":["Approximation effort should be concentrated in the later convolutional layers, which tolerate stronger multipliers with almost no accuracy loss.","Several configurations lose between 0.02% and 1% accuracy while still saving multiplication energy, so fine-grained approximation can be almost lossless on a quantized model.","Filter- and kernel-level allocation dominates the reported Pareto front, so layer-uniform approximation schemes leave energy savings unrealized.","Because the flow works on frozen quantized models without retraining, it remains usable when training data are proprietary or unavailable."],"supporting_citations":[{"why":"Supplies the ROUP approximate multiplier family, the central hardware object whose configurations are distributed across the network.","marker":"[2]"},{"why":"Defines the layer-wise approximation flow, the AxConv2D operator, and the no-retraining mechanism that MAx-DNN extends to finer granularities.","marker":"[7]"},{"why":"Provides the standard approximate-multiplier library used as the comparison baseline for the reported energy and accuracy advantages.","marker":"[5]"},{"why":"Supports the premise that DNN workloads are multiply-accumulate dominated, which grounds the multiplier-centric energy model.","marker":"[9]"}],"fun_headline_variants":["Per-kernel multiplier placement saves 54% DNN energy","Fine-grained inexact arithmetic cuts DNN energy by 54%","Kernel-level multipliers: 54% energy, 4% accuracy trade-off"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The energy model estimates each layer's consumption as the number of multiplications times the average energy of the assigned multiplier, omitting memory access, data movement, and control logic; if multiplier energy is not the dominant cost in a real accelerator, the reported 54% savings will not become system-level savings.","fun_headline_variants_meta":{"raw":{"variants":["Per-kernel multiplier placement saves 54% DNN energy","Fine-grained inexact arithmetic cuts DNN energy by 54%","Kernel-level multipliers: 54% energy, 4% accuracy trade-off"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000694,"raw_usage":{"total_tokens":3109,"prompt_tokens":885,"completion_tokens":2224,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":501,"completion_tokens_details":{"reasoning_tokens":2163}},"tokens_in":501,"tokens_out":2224,"duration_ms":23250,"temperature":1.0,"reasoning_tokens":2163,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T22:25:36.419375+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Build one of the reported Pareto-optimal approximate accelerators with a realistic memory hierarchy and measure total inference energy against the exact quantized accelerator; if multiplier energy is a small share of total system energy, the claimed 54% gain will not reproduce.","supporting_citations":[{"cited_title":"Cooperative Arithmetic-Aware Approximation Tech- niques for Energy-Efficient Multipliers,","cited_arxiv_id":null,"evidence_quote":"Supplies the ROUP approximate multiplier family, the central hardware object whose configurations are distributed across the network."},{"cited_title":"ALW ANN: Automatic Layer-Wise Approximation of Deep Neural Network Accelerators without Retraining,","cited_arxiv_id":null,"evidence_quote":"Defines the layer-wise approximation flow, the AxConv2D operator, and the no-retraining mechanism that MAx-DNN extends to finer granularities."},{"cited_title":"EvoApprox8b: Library of Approximate Adders and Multipliers for Circuit Design and Benchmarking of Approximation Meth- ods,","cited_arxiv_id":null,"evidence_quote":"Provides the standard approximate-multiplier library used as the comparison baseline for the reported energy and accuracy advantages."},{"cited_title":"Efficient AI System Design With Cross-Layer Approximate Computing,","cited_arxiv_id":null,"evidence_quote":"Supports the premise that DNN workloads are multiply-accumulate dominated, which grounds the multiplier-centric energy model."}],"review_version":1}