acceptodds
Under review as a conference paper at ICLR 2027

MUL-GPT: Timing-Structure-Aware Generative Optimization for Multiplier Circuits

Abstract

Multiplier optimization is a challenging combinatorial optimization problem, as its design space grows rapidly with increasing operand bit width. In recent years, learning-driven methods have provided a promising approach to discovering high-speed and area-efficient circuit structures. However, existing methods typically optimize only a small set of global post-synthesis metrics, such as area and delay, without explicitly modeling the internal timing structure of multipliers or the mechanisms underlying critical path formation. Consequently, they struggle to accurately identify timing bottlenecks and effectively exploit internal timing information to guide structural optimization. To address these limitations, we propose MUL-GPT, an internal-timing-structure-aware generative framework for multiplier optimization. MUL-GPT is a generative pretrained Transformer (GPT)-style generative framework that addresses the internally coupled timing optimization problem of multipliers. It unifies constrained autoregressive compressor-tree generation, arrival-time-conditioned adder optimization, and static timing analysis (STA)-guided local interconnection rewiring, with internal timing information serving as the link across these stages. Experiments show that the best designs found by MUL-GPT consistently outperform state-of-the-art methods across various arithmetic-unit design tasks, reducing the area-delay product (ADP) by up to 10.57%. These improvements are preserved when the optimized arithmetic units are integrated into large-scale artificial-intelligence (AI)-accelerator circuits, where MUL-GPT achieves an ADP reduction of up to 7.90%, further demonstrating its scalability and practical applicability.

Primary area:
infrastructure, software libraries, hardware, systems, etc.
Keywords:
Computing Circuits Optimization, Electronic Design Automation, Datapath Optimization, Reinforcement Learning

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_92QiOpenReview ↗
Rating
2(−1)
Confidence
2
Informativeness
3
Presentation
2(−1)
Soundness
3
Read review
Paper summary

This paper presents MUL‑GPT, an internal‑timing‑structure‑aware generative optimization framework for multiplier circuits. It reformulates compressor‑tree construction as constrained GPT‑style autoregressive sequence generation, adopts DAPO reinforcement learning with area‑delay‑product (ADP) derived from static‑timing‑analysis (STA) as reward, and alternately optimizes compressor‑tree topology and arrival‑time‑conditioned final‑adder generation. A critical‑path‑guided Cross‑Entropy‑Method (CEM) is further deployed for local interconnection rewiring. The framework outputs valid Verilog RTL. It achieves reduced ADP compared with existing state‑of‑the‑art approaches on 8‑to‑64‑bit AND‑encoded and Booth‑encoded multipliers as well as MAC units. Performance gains persist when optimized arithmetic components are integrated into AI‑accelerator PE arrays. Ablation studies verify the contribution of each component, and the local rewiring module is demonstrated to serve as a general‑purpose post‑processing plugin compatible with other multiplier optimization methods.

Critical strengths
  1. Joint Optimization of Time Sequence Awareness for the Complete Multiplier Chain: Unlike most methods that only look at area delay indicators globally, MUL-GPT explicitly models internal timing coupling relationships within the multiplier, jointly optimizing three stages: compressed tree generation, time-constrained adder generation, and local rerouting of critical paths; The adder adapts the non-uniform signals output from the compression tree to reach the cross-section, solving the pain point of "timing mismatch between the compression tree and the final adder," rather than independently optimizing the two submodules.
  2. Novel Structured Circuit Sequence Generation Solution: Transforms compression tree topologies into GPT autoregressive sequences with dynamic validity masks, using multi-branch cascaded prediction of column, level, and compressor types, filtering out invalid circuits during the generation phase to ensure the output structure meets hardware constraints; Paired with DAPO reinforcement learning, it rewards ADP from actual STA simulations using EDA tools, directly guiding model iteration with real comprehensive time-series feedback rather than relying on simplified theoretical models.
  3. Modular Reusable Critical Path Rerouting Component: Based on STA, the compressor slice corresponding to the critical path is positioned, and signal port redistribution is performed only in bottleneck regions using CEM cross-entropy methods, greatly reducing the search footprint; This rerouting module is not bound to the MUL-GPT master generator and can serve as a general-purpose post-processing plugin, directly enhancing the performance of multipliers produced by other algorithms such as Wallace, RL-MUL, ARITH-DAS, and more, making it highly versatile.
  4. Comprehensive and industrial-specific experimental verification: covers multipliers and MAC units encoded in 8-64 bit AND and Booth; Not only does it isolate evaluation arithmetic units, but it also integrates generative circuits into AI accelerator PE arrays and Koios 2.0 benchmarks, allowing validation unit-level optimization benefits to be retained in large-scale systems; Using Pareto Frontier + Hypervolume Metric for comprehensive area-delay trade-off, not only a single ADP scalar but also functional simulation and formal equivalence checks are performed to ensure the correctness of RTL generation.
  5. Detailed and controllable ablation and mechanism analysis: Through multiple ablation experiments, the compression tree generation, conditional adders, and CEM rewiring each contribute their own contributions; Conducting attention analysis confirmed that Transformer pays more attention to historical tokens strongly related to timing; Provides multi-turn independent operation of fluctuation statistics and training convergence curves; Fair comparison baseline, controlling candidate circuit search budgets to be consistent, excluding unfair comparisons caused by different search budgets.
Critical weaknesses
  1. Simulation evaluation relies on fixed 45nm NanGate process libraries, and generalization capability has not been fully verified. All experiments were based on the NanGate 45nm open-source unit library, with no control trials conducted at other process nodes (such as 7nm/28nm); Under different processes, the timing and area characteristics of compressor and adder units vary greatly, making it impossible to directly prove that the optimization benefits of this method can be seamlessly transferred to other processes.
  2. High computational overhead and high hardware resource threshold. A large amount of EDA synthesis and STA timing simulation is needed as reinforcement learning reward signals; 32-bit and 64-bit tasks must be trained using dual GPUs, and running them fully takes a long time. Although batch parallel sampling is used, it demands high computing power, EDA tool licenses, and CPU computing clusters, making it costly for ordinary researchers to replicate complete experiments.
  3. Only for compressed tree multiplier architectures, applicable scenarios are limited. The framework only supports compression tree multipliers composed of FA/HA, does not support other multiplication topologies, and does not extend to other arithmetic circuits such as division or root squares; The input is limited to powers of 2, a conventional bit-width multiplier, and does not evaluate asymmetric bit-width multiplication.
  4. Optimization goals heavily dependent on ADP scalars, Pareto has shortcomings in its optimization capability. In the reinforcement learning stage, 'Area × Delay' is directly used as a single-scalar reward, favoring a compromise solution; The training process does not explicitly maintain the complete Pareto solution, making it difficult to generate extreme solutions with extreme speeds or extremely small areas; A complete Pareto frontier can only be obtained through post-processing sampling, and the training phase lacks a multi-objective reinforcement learning mechanism.
  5. CEM Local Rerouting Assumptions Exist and Carry Failure Risk. The algorithm assumes that performance gains mainly come from critical path slicing, with port redistribution only on the critical path; When multiple paths approach critical paths (multiple bottlenecks), optimizing only a single set of key slices significantly reduces the yield. Rerouting only modifies the signal-port mapping and cannot change the type or position of the compressor itself; the optimization limit is constrained by the results generated by the pre-compression tree.
  6. Evaluate only at the level of logical synthesis without considering physical implementation effects. All metrics come from logical synthesis + STA, without place-route evaluation; Ignore routing delays, congestion, and routing area costs. RTL with excellent logical synthesis may see area-latency returns shrink or even reverse after backend layout and routing.
  7. 64-bit Equal Bit Width Exists Model Compromise. The 64-bit multiplier corresponds to a very long token sequence. To control memory usage, enabling sliding window attention while reducing the number of samples and iteration cycles is also reduced, thereby lowering the search budget. With large bit widths, model expression capability and search sufficiency are compromised.
  8. Lack of thorough noise and robustness testing. The experiment was conducted under ideal input conditions and did not study algorithm performance under perturbations such as defects in some input products or changes in timing constraints; Lack of robustness testing under different clock constraints and load conditions.
Questions for authors
  1. All experiments in this paper are based on the NanGate 45 nm process library, and the time-sequence-area characteristics of full and half adders differ significantly under different processes. Have you observed fluctuations in MUL-GPT's output structure as it migrates to other process nodes? What improvements will be made going forward so that the model can adapt to new processes without large-scale RL training?
  2. Currently, reinforcement learning uses ADP (area-delay product) as a single-scalar reward, and the Pareto solution is not explicitly maintained during training. Is there consideration for introducing multi-objective reinforcement learning, with direct sampling during training favoring extremely high-speed or ultra-small area candidates, rather than only post-processing to build Pareto frontiers?
  3. CEM rerouting only modifies the mapping between signal and port, without altering the compressor position or type; And only optimize critical path slicing. When the circuit has multiple near-critical paths (multiple bottleneck scenarios), the optimization benefit of this module decreases. Have you considered expanding the search scope for rewiring or dynamically identifying multiple candidate slices to further unlock optimization potential?
  4. All performance evaluations remain at the logic synthesis and STA levels, without layout or routing. RTL with excellent logical synthesis and backend routing introduces line delay and congestion that can affect actual performance. How do you plan to incorporate feedback from layout and routing into the optimization closed loop in the future? Is it possible to lightweightly estimate routing costs and avoid running the full physical implementation every time?
  5. 64-bit multipliers sacrifice some search sufficiency due to the long sequence tokens, use sliding window attention, and reduce search budget. For multipliers with larger bit widths, besides sliding windows, should we consider a strategy of layered/chunky compression tree generation, solving the memory and efficiency challenges caused by long sequences without significantly reducing search budgets?
5 comments
AuthorsOpenReview ↗

Response to Reviewer 92Qi

We sincerely thank you for your careful review and constructive, insightful, and valuable comments. We have uploaded the revised manuscript, with changes highlighted in blue. Below, we provide detailed responses to your questions.

W1, Q1—Process Adaptation and ASAP7 Evaluation

Thank you for your comments. The Nangate45 experiments alone do not establish that the optimization gains transfer directly to other processes. MUL-GPT’s structural representation and legality constraints are process-independent, while performance optimization relies on synthesis and timing feedback from the target process. A change in process therefore requires reoptimization. We will clarify the distinction between adapting the framework and transferring a trained model.

1.Cross-Library Resynthesis of Optimized RTL Designs

When RTL designs optimized under Nangate45 are directly resynthesized using ASAP7[1], the relative ADP rankings of the methods change. Synthesis-feedback-driven optimization may be more closely tailored to the source technology library, so its performance advantages may not transfer directly. Some methods relying on logic-level proxy models perform well in this transfer evaluation.

Method8-bit16-bit32-bit64-bit
MUL-GPT10.948767.6991366.34351836.1017
ARITH-DAS11.972067.7029355.53421771.8730
ArithTreeRL11.254768.2806357.72031821.4386
UFO-MAC10.578568.1711355.45721797.4971
GOMIL11.392268.2011356.64381681.3061
MUTE11.772467.2972381.94691841.4303
RL-MUL12.696379.1621385.07412013.5411
Wallace12.677777.2731409.93642054.3178

2.Reoptimization with the ASAP7 Library

However, when optimization is rerun using synthesis feedback from the target library, synthesis-feedback-driven methods often outperform those based on logic-level proxy models. Given the time constraints, we conducted additional 16- and 32-bit experiments at the ASAP7 FF corner with MUL-GPT and two state-of-the-art baselines, ARITH-DAS and ArithTreeRL, to assess the framework’s adaptability to a different process. MUL-GPT achieved lower ADP than both baselines at both bit widths.

Method16-bit32-bit
MUL-GPT54.4547301.2856
ARITH-DAS63.9996339.3232
ArithTreeRL63.3840352.7982

3.Future Work

Inspired by RL-Sizer [2], we plan to encode standard-cell area and port timing characteristics as conditioning inputs and train jointly across multiple processes. For a new process, we plan to reuse the existing model and explore calibrating a performance predictor with a small number of target-library synthesis samples, drawing on cross-node timing adaptation in AdaTimer[3]. We will also investigate parameter-efficient fine-tuning to reduce retraining costs.

W2—Computational Cost and Resource Requirements

Thank you for your comments. We agree that repeated logic synthesis and STA calls account for most of MUL-GPT’s computational cost, and that the full search takes considerable time at 32 and 64 bits. Batched sampling and parallel evaluation increase throughput, while the total computational cost remains. Our evaluation flow uses the open-source tools Yosys and OpenROAD and requires no commercial EDA licenses. Researchers with limited hardware can reduce the sampling batch size per iteration.

W3—Scope and Extensions to Other Circuits

Thank you for the suggestion. We focus on multiplier compressor trees built from full and half adders (FAs and HAs). We cover AND and Booth partial-product encodings and evaluate their practical use in MAC units and PE arrays. Our results on compressor and prefix trees suggest that autoregressive generation could extend to other circuits. Its applicability to other arithmetic circuits remains to be established.

We could support asymmetric operand widths by adjusting the initial partial-product distribution and additional compressor types, such as 4:2 compressors, by modifying the legality masks. Division and square root would require new task formulations. We will clarify these scope boundaries in the revision.

AuthorsOpenReview ↗

Response to Reviewer 92Qi

W4, Q2—Multi-Objective Optimization and Pareto Coverage

Thank you for the suggestion. The current ADP reward imposes a fixed area–delay preference and does not guarantee coverage of nondominated solutions under different preferences. A nondominated set obtained through post-processing also does not substitute for multi-objective exploration during training.

HAVE uses Pareto hypervolume gains to guide the selection of preferences and timing constraints. Adaptive Pareto Reset uses current nondominated designs and their perturbations to guide subsequent exploration. Inspired by these approaches, we plan to condition generation on optimization preferences, maintain an archive of nondominated designs during training, and adjust sampling based on frontier coverage, including extreme preferences that emphasize area or delay. We also plan to explore legal perturbations and continuations of compression sequences so that Pareto information can guide training directly, and to evaluate these changes under a matched synthesis budget.

W5, Q3—Rewiring Multiple Near-Critical Paths

Thank you for the suggestion. We agree that multiple near-critical paths can limit the gains from local rewiring. In an early experiment, we considered three paths simultaneously, but the larger search space yielded no clear additional gain under that configuration. We therefore optimize sequentially, rerunning STA after each accepted improvement to update the critical path and candidate slices. We plan to use timing slack to select multiple near-critical paths dynamically and group the search to control its cost. We will evaluate the gains under a matched synthesis budget.

W6—Post-Route Evaluation and Physical Feedback

Thank you for the suggestion. Post-synthesis area and STA delay only partially reflect the effects of interconnects, congestion, and physical optimization. We will clarify this limitation in the revision.

1.Post-Route

We use an OpenROAD placement-and-routing flow with Nangate45. We initialize the floorplan with 30% utilization, a 1.0 aspect ratio, and a 10 μm core boundary margin. We then insert tap cells, generate the power delivery network, and perform global placement, pin placement, detailed placement, global routing, and detailed routing. The target density for global placement is 0.3. After routing, we extract parasitics and check timing, area, and routing. The post-route ADP results are as follows.

Design8-bit16-bit32-bit64-bit
MUL-GPT218.351546.748948.6952794.37
ARITH-DAS252.311626.909938.1657000.39
ArithTreeRL247.061617.489618.6958429.97
UFO-MAC288.461935.6510965.2762621.39
GOMIL307.821949.5711029.7560513.83
MUTE257.421743.0610388.6856017.01
Wallace306.251926.7211156.6767251.55
RL-MUL267.701769.6410654.6158000.43

2.Future Work

We plan to use a staged evaluation. We will screen candidates through synthesis, then place and globally route a small set of promising designs. We will feed timing under estimated parasitics, wirelength, and congestion back into the optimization loop. We will also calibrate a lightweight cost model with physical implementation data to help evaluate the remaining candidates. The final selected designs will undergo full placement and routing to verify their performance while controlling evaluation cost.

AuthorsOpenReview ↗

Response to Reviewer 92Qi

W7, Q5—Scaling to Larger Bit Widths

We agree that hierarchical or blockwise generation is a promising direction for scaling to larger bit widths. MUL-GPT uses (column, compression stage, type) tokens to specify compressor choices explicitly. ArithTreeRL and RL-MUL select compressor types or counts by column and then assign stages and connections according to rules. Our finer-grained representation preserves more structural freedom. Even when the FA and HA counts are identical in each column, different stage assignments and connections can produce different propagation paths and arrival times. The tokens describe the structure, while wiring, synthesis, and STA determine the timing.

This representation can produce sequences of more than 4,000 tokens at 64 bits. A global planner could set compression targets for each block while local generators retain the operation-level representation. Blocks would pass carry weights, stages, and arrival times between them, and the complete design would be evaluated as a whole. This approach could shorten local sequences and reduce GPU memory requirements. Whether it can maintain optimization quality under the same search budget remains to be verified.

W8—Model Robustness

Thank you for the suggestion. Our experiments use ideal inputs and fixed evaluation conditions. Systematic perturbation and robustness tests remain to be conducted. The role of per-bit arrival times in guiding adder generation motivates us to include output loads and timing constraints as conditioning features for multiplier generation. We plan to explore training with multiple conditioning inputs and evaluate adaptation across load conditions, timing requirements, and input perturbations to reduce reliance on a single evaluation setup.

Reference

[1]Clark L T, Vashishtha V, Shifren L, et al. ASAP7: A 7-nm finFET predictive process design kit[J]. Microelectronics Journal, 2016, 53: 105-115.

[2]Lu Y C, Nath S, Khandelwal V, et al. Rl-sizer: Vlsi gate sizing for timing optimization using deep reinforcement learning[C]//2021 58th ACM/IEEE Design Automation Conference (DAC). IEEE, 2021: 733-738.

[3]Zhang X, Zhu B, Liu F, et al. Disentangle, align and generalize: Learning a timing predictor from different technology nodes[C]//Proceedings of the 61st ACM/IEEE Design Automation Conference. 2024: 1-6.

Reviewer_C561OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies the joint optimization of multiplier compression trees and final adders, and proposes MUL-GPT. It adopts GPT-style autoregressive generation for compression trees, fine-tunes with DAPO-style reinforcement learning, conditionally generates the final adder based on compression tree arrival times, and applies critical-path-driven CEM local rewiring for interconnect optimization. Experiments cover 8–64-bit AND/Booth multipliers, PE arrays, and Koios macro cell integration.

Critical strengths

For the reasonable scenario of EDA ML: multiplier compression tree + CPA co-optimization, a complete pipeline has been implemented, including generation -> STA -> rewiring -> alternating optimization and circuit structure generation, along with functional simulation and HimaFormal equivalence checking. The workload is quite substantial.

Critical weaknesses
  1. The "timing-structure-aware" aspect mentioned in the title lacks sufficient novelty. Existing work UFO-MAC has already identified that the CPA should adapt based on compression tree arrival times, which shares a similar motivation with this paper. This contradicts the claim in the abstract that "existing methods ... without explicitly modeling the internal timing structure of multipliers or the mechanisms underlying critical path formation." Moreover, the main body of the proposed method is a combination of existing techniques, with the primary contribution lying in system integration.
  2. Given that the contribution of this paper lies in system integration, it should include direct comparisons with methods such as UFO-MAC under the same synthesis flow and budget for 16/32-bit configurations. Complete ablation studies related to arrival conditioning should be provided, and the same search budget should be used in Table 1.
Questions for authors
  1. To ensure rigor, I suggest that the authors revise the phrasing in the abstract and introduction such as "existing methods ... without explicitly modeling the internal timing structure of multipliers or the mechanisms underlying critical path formation" to avoid overclaiming.
  2. Could the authors clearly clarify in the paper the essential differences from UFO-MAC, and provide a direct comparison with UFO-MAC under the same synthesis flow and budget for 16/32-bit configurations, explicitly specifying the differences in timing conditioning?
  3. Regarding the arrival conditioning results in Table 3, could the authors supplement an ablation study comparing a standalone adder generator versus the conditional version?
  4. The experiments in Table 1 should use the same search budget. If corresponding settings already exist, please specify them in the main text or appendix.
  5. The 64-bit experimental results appear to be relatively weaker. If this is the case, it should be clearly noted.
  6. The current experiments mainly target the 45nm node. Is support available for advanced process nodes, such as a 7nm PDK?
3 comments
AuthorsOpenReview ↗

Response to Reviewer C561

We sincerely thank you for your careful review and constructive, insightful, and valuable comments. We have uploaded the revised manuscript, with changes highlighted in blue. Below, we provide detailed responses to your questions.

W1, Q1—Clarifying Our Timing-Aware Contributions

Thank you for the correction. In the abstract and introduction, we will revise broad statements such as “existing methods do not explicitly model the internal timing structure” to “existing methods rely on heuristic rules and proxy delay models to characterize the internal timing structure.” We will explicitly acknowledge that methods such as UFO-MAC exploit nonuniform arrival times. We will also clarify our contribution: we use internal timing information to condition generation, learn topologies from synthesis feedback, and alternately optimize the compressor tree and final adder.

W1, Q2—Comparison with UFO-MAC

1.Methodological Differences from UFO-MAC

Thank you for the suggestion. UFO-MAC uses integer linear programming (ILP) to optimize compressor-stage assignments and interconnections. It chooses an initial adder structure based on the arrival-time profile, then estimates delay with a linear FDC model based on depth and fan-out to refine the prefix graph. MUL-GPT conditions prefix-structure generation on per-bit arrival times through cross-attention and evaluates candidates using the post-synthesis ADP of the full multiplier. Both methods use nonuniform arrival times, but they differ in how they search for structures and use timing information. UFO-MAC adjusts topology using a proxy timing model; MUL-GPT learns topology conditioned on arrival times through synthesis and STA feedback.

2.Performance Comparison with UFO-MAC

Table 1 compares the methods under a common final evaluation flow. A common synthesis flow does not imply a common search budget. UFO-MAC relies primarily on an ILP solver and analytical models, so synthesis counts alone do not capture its computational cost. MUL-GPT achieves substantially lower delay and ADP than UFO-MAC at every bit width.

TypeMethod8-bit16-bit32-bit64-bit
DelayADPDelayADPDelayADPDelayADP
ANDMUL-GPT0.6033215.98140.89511382.03441.20107700.81201.554840342.3956
ANDUFO-MAC0.7609286.85931.11351798.30251.42519290.22691.856847911.0104
BoothMUL-GPT0.6956255.28520.93931262.41921.23356189.70301.539829776.6524
BoothUFO-MAC0.8065307.27651.07491497.33571.52667745.96841.876535903.0745

W2, Q3—Standalone adder optimization

We add a standalone adder variant to the ablation study. We fix the prefix tree to the best adder obtained through independent ADP optimization under the same search budget, keeping all other settings unchanged. The results are as follows:

Bit WidthVariantArea (μm²)Delay (ns)ADP (μm²·ns)
16-bitStandalone Adder15430.98911526.1813
16-bitw/o Adder Opt.15380.93581439.2604
32-bitStandalone Adder64041.24577977.4628
32-bitw/o Adder Opt.64281.27258179.6300

The results for the independently optimized adders are shown below:

Adder widthMethodArea (μm²)Delay (ns)ADP (μm²·ns)
32-bitYosys behavioral adder2280.5138117.1464
32-bitStandalone optimized adder2230.395188.1073
64-bitYosys behavioral adder4770.6523311.1471
64-bitStandalone optimized adder4800.5002240.0960

Both achieve lower ADP than the corresponding Yosys behavioral adders. Notably, despite its lower standalone ADP, the optimized adder degrades 16-bit multiplier performance when replacing the Yosys behavioral adder. We hypothesize that late compressor-tree outputs traverse slow prefix-tree paths, compounding delay. This further motivates alternating optimization of both structures using internal timing information, with compressor-tree timing guiding adder generation.

AuthorsOpenReview ↗

Response to Reviewer C561

Q4—Search budget

Thank you for pointing out that our accounting of the search budget was incomplete. Tables 1, 2, and 9 and Appendix E.7 match only the budget for compressor-tree candidates. Conditional adder optimization and CEM rewiring are components of MUL-GPT whose performance gains come with additional computational cost. We will report these costs explicitly in the revision. Each MUL-GPT run is scheduled to evaluate 70×3×512=107,52070\times3\times512=107{,}520 compressor-tree candidates and an equal number of conditional adder candidates. Based on the observed average, CEM generates approximately 3×3×30×32=8,6403\times3\times30\times32=8{,}640 additional candidates, bringing the total to approximately 223,680 candidates per run. We conducted three independent MUL-GPT runs. To approximately match the compressor-tree search budget, each MUL-GPT run corresponds to 10 runs each of ARITH-DAS and MUTE at 10,000 compressor-tree candidates per run, 33 ArithTreeRL runs at 3,000 candidates per run, and three RL-MUL runs at 30,000 candidates per run. The search budgets are approximately halved for 64-bit designs. To distinguish the gains of the method from those of additional search resources, we also applied the same post-processing modules, with the same budget, to the 16- and 32-bit state-of-the-art baselines and compared performance under matched post-processing budgets. MUL-GPT remained ahead.

Q5—64-Bit Performance

Thank you for the suggestion. We will state more clearly that the 64-bit results are somewhat weaker and discuss how they relate to the resource settings. Given the much longer sequences, the 64-bit experiments use sliding-window attention with a window of 512 tokens, reduce the number of candidates per group from 512 to 256, and run a single round of tree and adder optimization with 200 iterations per stage in place of three alternating rounds. Model settings also affect these results, so we cannot attribute differences in gains across bit widths to scalability alone.

Q6—ASAP7 Evaluation

Thank you for the suggestion. Given the time constraints, we conducted additional 16- and 32-bit experiments at the ASAP7 FF[1] corner with MUL-GPT and two state-of-the-art baselines, ARITH-DAS and ArithTreeRL, to assess the framework’s adaptability to a different process. MUL-GPT achieved lower ADP than both baselines at both bit widths.

Method16-bit32-bit
MUL-GPT54.4547301.2856
ARITH-DAS63.9996339.3232
ArithTreeRL63.3840352.7982

Reference

[1]Clark L T, Vashishtha V, Shifren L, et al. ASAP7: A 7-nm finFET predictive process design kit[J]. Microelectronics Journal, 2016, 53: 105-115.

AuthorsOpenReview ↗

We Would Greatly Appreciate Your Feedback

Dear Reviewer C561,

We are writing as the authors of Submission 23581. We sincerely appreciate the time and effort you have devoted to reviewing our paper and thank you for your valuable comments.

As the author–reviewer discussion period is approaching its conclusion, we would be grateful if you could review our rebuttal and let us know whether our responses have adequately addressed your concerns. If so, we would greatly appreciate your acknowledgment. If any concerns remain, please let us know, and we will be happy to provide further clarification and respond promptly.

Thank you again for your insightful comments and valuable time. We look forward to your feedback.

Best,

The Authors of Submission 23581

Reviewer_rvDVOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The paper proposes MUL-GPT, a three-stage pipeline for multiplier optimization. In the first stage, a decoder-only Transformer generates the compressor tree as a sequence of (column, stage, type) tokens under dynamic legality masks. The model is pretrained on 10^6 random legal sequences with cross-entropy plus auxiliary legality and margin losses. It is then fine-tuned with DAPO-style group-relative policy optimization, using negative post-synthesis ADP as the reward, an elite replay pool, and KL regularization toward the pretrained policy. In the second stage, a generator built on the PrefixGPT backbone produces the final prefix adder, conditioned through cross-attention on the per-bit arrival-time profile of the compressor-tree outputs. The first two stages alternate for three rounds. In the third stage, a Cross-Entropy Method with Gumbel perturbation and Hungarian matching reassigns source signals to compressor input ports, restricted to the compressor slices traversed by the STA critical path. Evaluation uses Yosys, ABC and OpenROAD STA on NanGate 45nm. It covers AND and Booth multipliers from 8 to 64 bits, MACs, 8x8 PE arrays, and six Koios 2.0 designs. The headline claims are up to 10.57% ADP reduction over the best prior method on multipliers and up to 7.90% on PE arrays.

Critical strengths

S1. The problem is well motivated.

S2. The proposed method consistently beats the baselines.

S3. Provides the necessary material to reproduce the results.

Critical weaknesses

W1. The search budget is not matched. Section E.7 counts only compressor-tree candidates toward the matched budget. The conditional adder stage adds roughly another 107,520 full multiplier syntheses per run (70 iterations × 512 candidates × 3 rounds), and CEM rewiring adds 32 samples × 30 rounds per critical-path re-optimization. Table 11 implies about 3,000 candidates per ArithTreeRL run and about 10,000 per MUTE or ARITH-DAS run, against about 107,000 per MUL-GPT run. The paper never states how many baseline runs were performed or whether Table 1 reports the best over them. The per-1,000-candidate runtime for MUL-GPT appears to divide total wall-clock time, including the adder and CEM stages, by the tree count alone.

W2. The gain attributable to the generator is small. In Table 10, ARITH-DAS with the same conditional adder and rewiring reaches 1415.35 at 16 bits and 7908.90 at 32 bits. Against MUL-GPT’s means from Table 5 (1387.57 and 7785.74), the gaps are 1.96% and 1.56%. This margin is obtained with roughly twice the synthesis budget and three rounds of alternation, against a single post-hoc pass for the baseline. The “w/o fine-tuning” variant, which is essentially a learned sampler of random legal trees, reaches 1546.32 at 16 bits. That already beats RL-MUL (1701.73) and MUTE (1622.87), which suggests that much of the headline margin comes from the downstream modules and the larger budget.

W3. Several ablation effects are within run-to-run noise. Table 5 gives a standard deviation of 79.95 at 32 bits, which is 1.03% of the mean. Removing rewiring changes 32-bit ADP by 0.63%, and removing arrival conditioning changes it by 1.20%. The paper does not say whether the ablations are single runs or best of three. On the strongest baselines in Table 6 (MUTE, ArithTreeRL, ARITH-DAS), rewiring yields only 0.12% to 2.05%, so the claim of a “consistent gain” is not supported at 32 bits.

W4. The rewiring module closely follows DOMAC (Xue et al., arXiv 2503.23943), which is not cited. DOMAC represents the source-to-port mapping at each column and stage as a bijection matrix, relaxes it to a doubly stochastic matrix, and legalizes it with the Hungarian algorithm. MUL-GPT keeps this decision variable and this legalization step. It adds Gumbel perturbation, which is the Gumbel-Matching distribution of Mena et al. (ICLR 2018, also uncited), and replaces gradient descent with Rubinstein’s CEM update. Only the critical-path restriction is the authors’ own choice.

W5. The novelty of the compressor-tree generator is overstated. PrefixGPT is cited only as the backbone of the adder stage. The compressor-tree generator follows the same recipe: coordinate tokens, a legality mask, pretraining on random valid designs, and RL fine-tuning. Contribution (1) nonetheless presents this formulation as new. The genuinely new parts, the cascaded column, stage and type heads and the legality and margin losses, are not ablated against a plain PrefixGPT-style setup. Learned co-optimization of the compressor tree and the adder was also explored by GOMARL (Feng and Wang, ISEDA 2024), which is not cited.

W6. The positioning on final adders is inaccurate. Limitation (2) in Section 3 states that existing methods lack an explicit mechanism for adapting the adder topology to non-uniform arrival times. This is contradicted by UFO-MAC and by Ramkumar and Kittur, both of which the paper cites.

W7. The evaluation flow is too weak to support interconnect-level claims. Synthesis runs flatten followed by “abc -fast”, which can restructure the RTL port assignment that the rewiring module optimizes. Results are post-synthesis only, with no placement or routing, even though OpenROAD is already in the toolchain.

Questions for authors

Q1. For each baseline in Tables 1, 2 and 7, how many independent runs did you perform, how many synthesized candidates did each run evaluate, and is the reported value the best over all runs? Please also report the total number of synthesis and STA calls per method. For MUL-GPT, include the conditional adder candidates and the CEM rewiring samples.

Q2. Table 11 normalizes MUL-GPT’s runtime to 1,000 explored candidates. Does that figure divide the total wall-clock time, including the adder and CEM stages, by the number of compressor trees alone? How many parallel EDA workers did MUL-GPT use, and how many did each baseline use?

Q3. Can you report a comparison in which ARITH-DAS and ArithTreeRL receive the same total synthesis budget as MUL-GPT, all methods use the same final adder, and all methods receive the same three rounds of tree and adder alternation that MUL-GPT receives?

Q4. Which final adder does each baseline use in Table 1? If it differs from what MUL-GPT uses, please report all methods under one fixed adder, for example the Yosys behavioral adder or a fixed Sklansky adder.

Q5. The “w/o fine-tuning” variant reaches 1546.32 at 16 bits, which is better than RL-MUL and MUTE. How does a plain random legal sampler perform with the same budget and the same downstream adder and rewiring? This would show whether pretraining adds anything beyond the legality masks.

Q6. Are the entries in Table 3 single runs or best of three? Please report mean and standard deviation over at least three seeds for every ablation. At 32 bits, the effects of removing rewiring (0.63%) and arrival conditioning (1.20%) are close to or below the run-to-run standard deviation in Table 5 (1.03%).

Q7. Your rewiring module shares its per-(column, stage) bijection matrix and its Hungarian legalization with DOMAC (Xue et al., arXiv 2503.23943). ARITH-DAS, whose setup you otherwise adopt, includes DOMAC as a baseline. Why was DOMAC omitted? Please compare against it directly under identical compressor trees and synthesis budget.

Q8. What is the “delay-based heuristic” used for the initial port assignment in Appendix B.4? If it is the three-greedy rule of TDM (Oklobdzija et al., IEEE TC 1996), please cite it. Please also report the gain of CEM over that initial assignment alone, so that the benefit of CEM can be separated from that of any timing-aware assignment.

Q9. The compressor-tree generator follows the PrefixGPT recipe: coordinate tokens, a legality mask, pretraining on random valid designs, and RL fine-tuning. Apart from the change of token alphabet, what do you consider new relative to PrefixGPT? Can you ablate the cascaded column, stage and type heads and the legality and margin losses against a plain PrefixGPT-style decoder?

Q10. How does your alternating optimization of the compressor tree and the adder differ from GOMARL (Feng and Wang, ISEDA 2024), which co-optimizes the two stages with cooperating RL agents?

Q11. Your synthesis script adds flatten, opt and “abc -fast” to the ARITH-DAS script. Why was this change made, and how do the rankings in Table 1 change under the original script? How much of the rewired port assignment survives flattening and ABC restructuring? For example, what fraction of rewired connections remain traceable in the mapped netlist?

Q12. Can you report post-placement-and-routing area and delay from OpenROAD for the main 16-bit and 32-bit designs, along with results on a second library such as ASAP7?

6 comments
AuthorsOpenReview ↗

Response to Reviewer rvDV

We sincerely thank you for your careful review and constructive, insightful, and valuable comments. We have uploaded the revised manuscript, with changes highlighted in blue. Below, we provide detailed responses to your questions.

W1, Q1, Q2 — Search budget

  1. Detailed Search Configuration

Thank you for pointing out that our accounting of the search budget was incomplete. Tables 1, 2, and 9 and Appendix E.7 match only the budget for compressor-tree candidates. Conditional adder optimization and CEM rewiring are components of MUL-GPT whose performance gains come with additional computational cost. We will report these costs explicitly in the revision. Each MUL-GPT run is scheduled to evaluate 70×3×512=107,52070\times3\times512=107{,}520 compressor-tree candidates and an equal number of conditional adder candidates. Based on the observed average, CEM generates approximately 3×3×30×32=8,6403\times3\times30\times32=8{,}640 additional candidates, bringing the total to approximately 223,680 candidates per run. We conducted three independent MUL-GPT runs. To approximately match the compressor-tree search budget, each MUL-GPT run corresponds to 10 runs each of ARITH-DAS and MUTE at 10,000 compressor-tree candidates per run, 33 ArithTreeRL runs at 3,000 candidates per run, and three RL-MUL runs at 30,000 candidates per run. The search budgets are approximately halved for 64-bit designs. To distinguish the gains of the method from those of additional search resources, we also applied the same post-processing modules, with the same budget, to the 16- and 32-bit state-of-the-art baselines and compared performance under matched post-processing budgets. MUL-GPT remained ahead.

  1. Runtime Normalization and EDA Parallelism

To align with the search budget accounting above, Table 13 divides the end-to-end runtime, including conditional adder optimization and CEM rewiring, by the compressor-tree candidate count and reports the result per 1,000 candidates. MUL-GPT has a larger total search budget per run, but batched sampling during GRPO enables parallel evaluation with 64 EDA workers. The baselines generate fewer candidates per update. ARITH-DAS and MUTE use two EDA workers each, while ArithTreeRL and RL-MUL use one each. MUL-GPT finds better designs at a manageable runtime cost.

W2, Q5, Q3 — Gains attributable to the generator

  1. Post-Processing Budget Clarification

The synthesis budget for baseline post-processing is reported in the Appendix E.7 and is comparable to that for three rounds of alternating optimization. The 16-bit and 32-bit ARITH-DAS results of 1415.35 and 7908.90 are best values and should be compared with the corresponding MUL-GPT best values of 1382.0344 and 7700.8120, which represent reductions of 2.35% and 2.63%, respectively.

  1. Comparison with a Random Legal Sampler

Under the same compressor-tree search budget, randomly generated multipliers achieved best ADPs of 1746.088 at 16 bits and 9933.35 at 32 bits, comparable to Wallace multipliers. This highlights the importance of reinforcement learning fine-tuning.

  1. Comparison with Matched Budgets and Alternating Optimization

Under the same budget and three-round alternating schedule, ARITH-DAS and ArithTreeRL achieved the results below. These results were mostly worse than those reported in the Appendix E.7. MUL-GPT alternately updates the compressor tree and conditional adder using full-multiplier synthesis feedback. For ARITH-DAS and ArithTreeRL, changing the compressor tree and its arrival-time profile every 70 adder iterations may increase the cost of adaptation. Keeping the compressor tree fixed while optimizing the adder for 210 consecutive iterations provides a more stable arrival-time profile and can reduce repeated adaptation. The benefit of the optimization schedule therefore depends on the search mechanism.

Method16-bit32-bit
ARITH-DAS + alternating optimization1463.04737974.9924
ArithTreeRL + alternating optimization1450.81268248.3968

AuthorsOpenReview ↗

Response to Reviewer rvDV

W3, Q6—Multi-Seed Ablation Results

Thank you for pointing this out. The ablation experiments in Table 3 use the seed corresponding to the best-performing full-model run. We will clarify this in the revised manuscript. Given the time constraints, we prioritized three-seed experiments for the two ablations with smaller performance gaps: removing arrival conditioning and removing rewiring. The mean ± standard deviation of ADP is reported below:

Method16-bit32-bit
Full1387.5684 ± 7.40967785.7406 ± 79.9516
w/o Arrival Conditioning1404.2925 ± 9.18147835.6324 ± 61.0801
w/o Rewire1420.5374 ± 25.66047855.0153 ± 107.3869

The full model achieves lower mean ADP at both bit widths, although the 32-bit results exhibit greater variability across seeds. In the revision, we will replace “ consistent gain” in the discussion of interconnect rewiring with “a modest ADP reduction in the reported configurations” and interpret the small differences at 32 bits with particular caution.

W4, Q7—Relation to DOMAC

1.Differences from DOMAC

Thank you for pointing out the missing citations. We will cite DOMAC[1] and Mena et al. [2] in the revision and clarify how their work relates to ours. We draw on DOMAC’s representation of compressor-tree connectivity. DOMAC relaxes discrete wiring assignments into continuous variables and optimizes them by gradient descent through differentiable timing and area models. MUL-GPT searches discrete assignments within a small set of slices associated with the critical path and iteratively selects candidates to improve connectivity.

2.DOMAC Results

DOMAC does not provide a public implementation, and its results reported in ARITH-DAS fall short of those of the stronger baselines. Our reimplementation likewise performed relatively poorly, with ADPs of 1718.25 and 9855.00 at 16 and 32 bits, respectively. The fidelity of our reimplementation is difficult to verify.

W5, Q9, Q10—Distinctions from PrefixGPT and GOMARL

1.Differences from PrefixGPT

Thank you for your comments. MUL-GPT’s novelty lies in formulating compressor-tree optimization as constrained autoregressive sequence generation and designing a generation mechanism around the dependencies among compression operations. Compressor trees and prefix trees require different representations and legality masks. PrefixGPT represents prefix trees with node coordinates, whereas each MUL-GPT token represents a compression operation defined by its bit column, stage, and compressor type. Each operation changes the available signals. The legal stages therefore depend on the selected column, and the legal compressor types depend on both the selected column and stage.

2.Rationale for the Three Cascaded Prediction Heads

To model these dependencies, we extend the two-head architecture to a column → stage → type cascade. The prediction heads pass conditioning features to subsequent heads, dynamic masks enforce operation legality, and auxiliary losses help the model distinguish legal from illegal choices before masking. Together, the three heads specify a complete compression operation. Removing any one head leaves the compression decision underspecified.

3.Differences from GOMARL

GOMARL[3] jointly executes the compressor-tree and adder agents’ actions at each environment step and trains the agents using a shared full-multiplier reward and a joint Q-value. MUL-GPT alternates the searches by round. It first fixes the adder and optimizes the compressor tree, then conditions prefix-adder generation on the per-bit arrival times extracted by STA to match the resulting tree.

W6—Clarifying Prior Timing-Aware Methods

Thank you for the correction. Our statement that prior methods lack a mechanism for adapting to nonuniform arrival times was too broad. We will revise Limitation (2) to read “Existing methods adapt final adders to nonuniform arrivals using simple heuristics and proxy delay models.”

1.Differences from UFO-MAC

Ramkumar and Kittur already combine different adder structures according to the arrival-time profile. UFO-MAC builds on this approach by estimating per-bit delays with a linear FDC model that accounts for depth and fan-out, then using these estimates to adjust the depth and fan-out of the initial prefix graph. These methods rely on heuristic rules and proxy delay models to characterize internal timing. MUL-GPT directly conditions prefix-tree generation on arrival times through cross-attention and learns topology from synthesis feedback.

AuthorsOpenReview ↗

Response to Reviewer rvDV

W7, Q11—Synthesis Flow

Thank you for raising the effect of the synthesis flow on the results. Our flow follows the ArithTreeRL setup. The flatten pass flattens the module hierarchy, opt simplifies the logic, and abc -fast performs fast technology mapping. These passes tend to reduce area and delay. In practice, small arithmetic units such as multipliers are typically flattened before standard-cell mapping. Rankings vary under the ARITH-DAS script because the designs in Table 1 were optimized under a different flow. This experiment shows how existing RTL performs across synthesis flows. Assessing the relative performance of the optimization methods under the new flow would require reoptimizing the designs under that flow. Nevertheless, MUL-GPT still achieves competitive ADP results under the ARITH-DAS flow.

design8-bit16-bit32-bit64-bit
MUL-GPT416.432491.3412936.6367845.48
ARITH-DAS406.452768.4820705.75142241.32
ArithTreeRL400.762513.3313170.864598.13
UFO-MAC342.832780.7114251.6782651.48
GOMIL361.742683.2619546.84127817.81
MUTE388.12632.6718819.15123265.63
Wallace418.313109.0722487.58153994.69
RL-MUL447.263145.2422600.39154652.15

Approximately 60–96% of rewired connections remain traceable in the mapped netlist. Two cases are difficult to trace: swapping only a and b of a full adder while keeping c fixed, and swapping a and b of a half adder. The two ports are logically symmetric, so the mapped gates still receive the same pair of signals after the swap. We can recover the signal pair, but the mapping from each signal to RTL port a or b is ambiguous.

Q4—Comparison with a Fixed Behavioral Adder

MUL-GPT uses its own adder generator, ArithTreeRL optimizes the adder with MCTS, and UFO-MAC uses FDC to optimize the adder. The remaining methods use the Yosys behavioral adder. The following are the additional 16-bit and 32-bit best results with the adder fixed to the Yosys behavioral implementation.

VariantAreaDelayADP
UFO-MAC 16-bit AND1549 μm²0.9886 ns1531.34 μm²·ns
UFO-MAC 32-bit AND6280 μm²1.4722 ns9245.42 μm²·ns
ArithTreeRL 16-bit AND1552 μm²1.0069ns1562.71μm²·ns
ArithTreeRL 32-bit AND6448 μm²1.3545 ns8733.82μm²·ns
MUL-GPT 16-bit AND1538 μm²0.9358ns1439.26μm²·ns
MUL-GPT 32-bit AND6395 μm²1.2531 ns8013.57μm²·ns

Q8 — CEM and port assignment

Thank you for the suggestion. We will cite TDM in the revision. We use a timing-aware assignment heuristic that gives later-arriving signals priority for faster input ports, consistent with the port assignment principle in TDM[4]. The same principle is widely used in Mullet multiplier generators, UFO-MAC, DOMAC, and ARITH-DAS.

The compressor-tree connections before rewiring already follow this principle. Appendix C.3 therefore shows the additional benefit of rewiring a single critical-path slice over timing-aware assignment alone, with a mean ADP reduction of 27.65 on randomly generated 16-bit multipliers. The benefits of CEM rewiring are further examined through the ablation study in Table 3 and the before-and-after comparisons in Table 8. The reported ADP reductions vary across designs, with larger improvements observed for some weaker designs and modest gains for stronger baselines.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.