Pith. sign in

REVIEW 4 major objections 6 minor 43 references

Iceberg: Enhancing HLS Modeling with Synthetic Data

T0 review · 4 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read Synthetic data cuts chip-design latency-model error by 86 percent.

desk verdict Solid, well-engineered synthetic-data recipe for HLS modeling, but the headline 86.4% gain rests on six noisy benchmarks and needs statistics before it should be quoted. read the letter →

arxiv 2507.09948 v2 pith:ABMIHYAW submitted 2025-07-14 cs.LG cs.AR

Zijian Ding , Tung Nguyen , Weikai Li , Aditya Grover , Yizhou Sun , Jason Cong This is my paper
classification cs.LGcs.AR
keywords high-levelsynthesissyntheticdataaugmentationfew-shotlearningTransformerNeuralProcessweaklabelsgraphnetworksLLMprogramgenerationdesignspaceexploration
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper argues that the generalization gap in learned high-level synthesis (HLS) performance models can be closed by synthesizing both sides of the training distribution: new C programs from an LLM and cheap weak labels for unmeasured design configurations. The authors introduce Iceberg, a dataset of thousands of LLM-generated synthesizable programs with actual latency labels on a subset, plus a weak-label generator built from an ensemble of GNNs with dropout. To learn from the hybrid data, they pair a GNN encoder with a Transformer Neural Process (G-TNP), an in-context model that adapts at test time from a handful of labeled examples. Across six real-world benchmarks, pretraining plus this weak-label augmented in-context model reduces geometric-mean squared error by 86.4% relative to the prior state-of-the-art baseline and brings offline best@1 design-space choices close to the ideal configuration. If the result holds, scarce HLS training data is not an obstacle to accurate few-shot performance prediction.

What carries the argument

The load-bearing pieces are: an LLM program generator that enforces tool-specific formatting and diversity constraints through iterative zero-shot chain-of-thought feedback; a weak-label generator that samples an ensemble of GNNs trained on each program's actual labels, applies random dropout to create a family of synthetic performance functions, and evaluates them on unseen configurations; and G-TNP, a Transformer Neural Process whose context and target tokens are GNN embeddings of the program's graph, enabling in-context meta-learning over (program, function) pairs. Weak labels are effective only when they come from multiple diverse synthetic functions rather than a single function.

What would settle it

Run the same few-shot adaptation on additional real-world programs outside the six tested, comparing Ice-H against Ice-A and against H-MoE; if the geometric-mean MSE improvement over H-MoE drops toward zero, or weak-label training stops beating actual-label-only training, the central claim fails. A cheaper check is to measure the ensemble GNN's test MSE on unseen configurations of new synthetic programs: values near or above the reported 0.321 on Iceberg would predict that failure.

Watch

Extended reading notes

Core claim

The central claim is that scaling synthetic data along both axes of an HLS design space—programs and pragma configurations—transfers to real programs. Using an LLM prompt with iterative chain-of-thought feedback, the authors generate 4,449 C programs compatible with Vitis HLS, run AutoDSE on 214 of them to get 14,840 actual labels, and then use an ensemble of GNNs trained on those labels, with MC-dropout to cheaply diversify the ensemble, to produce weak labels for unseen design configurations. A G-TNP model trained on the hybrid set learns a distribution over performance functions, so at adaptation time it can condition on a few actual labels of a new program and predict latencies across its design space. The paper reports that this recipe beats the Hierarchical-MoE state of the art on the HLSyn test programs by 31.5–37.9% after fine-tuning and by 86.4% (geometric-mean MSE) on six real-world applications, with best@1 latencies near the ideal.

Load-bearing premise

The load-bearing premise is that the GNN ensemble's weak labels are a sufficiently accurate proxy for true HLS latency on unseen design configurations; the paper itself reports this proxy is much weaker on the Iceberg set (average test MSE 0.321) than on HLSyn (0.047), so if it degrades further on broader distributions the headline gains could shrink or vanish.

Editorial extensions

If this is right

  • Pretraining on the Iceberg dataset improves generalization for model architectures it was not designed for: HARP's test accuracy rises by 80% and H-MoE's by 10% compared with training from scratch.
  • With the G-TNP architecture, fine-tuning on a few-shot adaptation set outperforms H-MoE on the ten HLSyn test programs by 31.5% (actual labels only) and 37.9% (hybrid labels).
  • On the six real-world applications, the hybrid-label model Ice-H beats H-MoE's geometric-mean MSE by 86.4% even without any fine-tuning.
  • Weak-label augmentation is most valuable when actual labels are scarce: on the small HLSyn training set Ice-H's MSE is 55% lower than Ice-A's, whereas on the larger Iceberg set the gap narrows.
  • In offline design-space optimization, the fine-tuned Iceberg models reach best@1 latencies within a small factor of the ideal point and outperform H-MoE by about 2.47x on HLSyn and 1.12x on the real-world set.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • If weak-label quality is the bottleneck, the paper's own numbers suggest a testable priority: spend marginal compute on actual labels rather than more weak labels whenever the GNN proxy is inaccurate, e.g. at the reported average test MSE of 0.321 on the Iceberg set.
  • The same recipe—tool-constrained LLM program generation, ensemble weak labels, and an in-context encoder—should transfer to other accelerator DSLs or toolchains that impose strict syntactic constraints, since the prompt is the main tool-specific component.
  • The released 3,401 synthesizable programs without labels are a resource the paper does not exploit; a semi-supervised variant that labels them with the GNN before training could test whether program diversity, not label count, drives the gain.
  • Because the 86.4% figure is a geometric mean over six programs where the baselines fail badly (for example conv2d), the headline improvement depends on the test distribution; a broader benchmark with more programs would show how robust the gain is.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 6 minor

Summary. The paper proposes Iceberg, a synthetic data augmentation pipeline for HLS performance modeling. It expands the program dimension with LLM-generated synthesizable C programs, expands the design-configuration dimension with weak labels produced by an ensemble of GNNs plus MC-dropout, and trains a Transformer Neural Process with a GNN encoder (G-TNP) on actual and weak labels. The evaluation compares Iceberg variants against HARP and Hierarchical-MoE on the HLSyn benchmark and on six real-world programs, reporting an 86.4% geometric-mean MSE improvement over H-MoE on the real-world few-shot adaptation task and improved offline DSE best@1 results. Code and data are said to be open-sourced.

Significance. If the headline results hold, the paper makes a useful step toward generalizable HLS performance models by combining synthetic pretraining with in-context meta-learning, addressing a real labeled-data scarcity problem. The evaluation has notable strengths: held-out real HLS labels, three random seeds for prediction, and ablations that isolate the pretraining dataset, the weak labels, and the G-TNP architecture. The paper is also honest about the weak-label accuracy limitation (Table IV) and its effect on the Ice-A versus Ice-H gap (Fig. 7b). The main weakness is that the central 86.4% claim rests on six real-world programs with no uncertainty quantification, and the LLM-based synthetic program generation is not described at the level needed for reproducibility. The weak-label concern raised in prior review is valid but secondary, because Ice-A already achieves most of the reported improvement; the statistical fragility of the six-program comparison is the load-bearing issue.

major comments (4)
  1. [§IV-H, Table IX, Fig. 5b] The headline 86.4% geometric-mean MSE improvement over H-MoE rests entirely on six real-world programs, yet no confidence interval, paired test, or leave-one-out analysis is reported. The programs are not independent: att-3mm and att-3mm-fuse are structural variants, and the per-program variance is high (e.g., conv2d Ice-H is 1.06±0.89, and optical-flow favors H-MoE at 0.32 vs 0.39). The entry att-3mm-fuse = 0.00±0.00 is rounded to two decimals and, as the smallest value in the geomean, has disproportionate influence; the exact value must be reported. I request a paired bootstrap or permutation test over programs, exact geomeans with confidence intervals, and a leave-one-program-out analysis to show that the reported improvement is not driven by one or two benchmarks.
  2. [§II (Synthetic P generation)] The LLM-based synthetic program generator is a core component of the method, but the paper does not disclose the LLM model or version, the initial prompt template, the iterative generation protocol, or the criteria used to select the 214 programs with actual labels out of 3401 synthesizable programs. A repository link alone is insufficient unless the prompts, generation logs, and full program set are actually released. Without this information the synthetic-data contribution cannot be reproduced or independently assessed for diversity and validity. Please specify the model and release the prompts and artifacts.
  3. [§IV-C, Table IV, Fig. 7b] The paper's own Table IV shows that the weak-label generator is much less accurate on the Iceberg pretraining set than on HLSyn (average test MSE 0.321 vs 0.047), and Fig. 7b shows the Ice-A versus Ice-H gap shrinks when pretraining on Iceberg. In Table IX, the headline 86.4% is reported for Ice-H, while Ice-A already achieves an 81.5% geomean reduction; the paper should quantify the marginal contribution of weak labels on the real-world set and its uncertainty. As written, the weak-label contribution is not cleanly separable from the pretrained G-TNP architecture, and the fixed 0.5 weak-to-actual ratio is not varied.
  4. [§IV-I, Fig. 6, Tables X-XI] The offline DSE metric is not fully defined: the figure of merit underlying the claimed 2.47× and 1.12× improvements is not stated explicitly (presumably geomean best@1 latency), and the treatment of invalid selected designs is inconsistent across rows (e.g., optical-flow is listed as 'Resource Over Utilization' in Table XI). The paper should state the metric, explain how invalid selections enter the reported geomean, and report the same uncertainty analysis requested for the prediction results, since the DSE claim inherits the fragility of the small benchmark.
minor comments (6)
  1. [§IV-F and §III-B] There are typos in the manuscript: 'risises' in §IV-F and 'necessities' in §III-B should be 'rises' and 'necessitates'.
  2. [Appx. C, Fig. 9] The caption text is contradictory: the first sentence says the blue curve indicates the Gaussian Process baseline, while two paragraphs later the blue curve is said to represent Iceberg's weak-label approach. Please clarify which curve corresponds to which method.
  3. [Abstract, §IV-H, Tables X and XI] The abstract's '86.4%', '2.47×', and '1.12×' numbers should be reconciled with the exact values in the tables and reported with rounding conventions; the current Table IX reports 0.13/0.92 = 85.9% rather than 86.4%, and the DSE factors should be computed from the geomeans shown in Tables X and XI.
  4. [Table VI vs Table V] Table VI reports H-MoE from-scratch test MSE as 0.12, while Table V reports H-MoE geomean as 0.08 on a six-program subset; the relationship between these two evaluation sets and protocols should be clarified.
  5. [Table IX] The att-3mm-fuse entry reported as 0.00 (0.00) should be given with more significant digits; a rounded zero in a geometric-mean calculation is information-lossy and materially affects the headline number.
  6. [Abstract] The GitHub URL contains an apparent typo ('UCLA-V AST'); please correct the link.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the headline 86.4% improvement is measured on held-out real HLS labels, and the weak labels are pseudo-labels used only as training targets, not as evaluation labels.

full rationale

The derivation chain is self-contained relative to the headline claim. Iceberg's weak labels (Section III-C) are produced by GNN ensembles fitted to actual labels of the training programs, but they are used only as training targets for the G-TNP model. The evaluation of the 86.4% improvement (Section IV-H, Fig. 5b, Table IX) is on held-out real-world programs from Rosetta and the attention layer, with actual Vitis HLS labels, so the result is not forced by construction. The few-shot context or fine-tuning points are sampled from the held-out real labels, not from the weak-label set. The use of HARP's encoder and of HARP/H-MoE as baselines involves same-group prior publications, but these serve as an architectural component and comparison baselines; they do not supply the target result. The paper also explicitly discloses the weak-label fidelity limitation (Table IV and Appx. Fig. 7b), which is a correctness caveat rather than a circular reduction. No fitted parameter is renamed as a prediction of the headline quantity.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

The central claim rests on assumptions about the transferability of LLM-generated programs, the quality of GNN weak labels, the validity of in-context adaptation, and the correctness of external resource filters. No new physical entities are introduced.

free parameters (4)
  • weak-to-actual label ratio = 0.5
    Hyperparameter controlling the mix of weak and actual labels in the hybrid training set (Sec. IV-C). Not tuned across datasets; the paper notes the best ratio is not explored.
  • number of (Program, Function) pairs L = 300
    Number of weak-label contexts sampled for pretraining (Table III caption). Chosen to balance compute and coverage.
  • number of configurations per pair K = 100
    Number of unseen design configurations sampled per (P, theta) pair (Table III caption).
  • GNN ensemble size = 19 (HLSyn), 14 (Iceberg)
    Number of GNNs in the ensemble for weak label generation (Table IV). Varies by dataset; the paper does not study sensitivity.
assumptions (5)
  • domain assumption LLM-generated programs are compatible with HLS tools and diverse enough to span the distribution of real-world kernels.
    Sec. II assumes the LLM with iterative feedback produces programs that transfer to benchmarks like Rosetta and attention layers; no direct measurement of program distribution overlap is provided.
  • domain assumption GNN ensembles trained on actual labels yield weak labels that are well-calibrated proxies for the true HLS performance function.
    Sec. III-C and Table IV; the paper acknowledges this assumption degrades on the Iceberg dataset (MSE 0.321), weakening the weak-label benefit.
  • domain assumption In-context adaptation with few-shot examples is sufficient for the G-TNP model to transfer from synthetic pretraining to a new program.
    Sec. III-B and Sec. IV-G/H; the model's ability to generalize is demonstrated but not proven for arbitrary programs.
  • domain assumption The resource and classification filters from Hierarchical-MoE correctly identify valid designs in the DSE evaluation.
    Sec. IV-D uses H-MoE's resource/classification models to filter designs before applying the Iceberg performance model; errors in these filters would affect the best@1 results.
  • domain assumption Synthetic programs generated by the LLM are not near-duplicates of the real-world test programs.
    Sec. II does not report any decontamination check against the HLSyn or Rosetta test programs; if overlap exists, the few-shot adaptation advantage could be inflated.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Iceberg: Enhancing HLS Modeling with Synthetic Data." pith.science (2026). https://pith.science/paper/ABMIHYAW

@misc{pith2026250709948,
  author       = {Pith},
  title        = {Pith review of: Iceberg: Enhancing HLS Modeling with Synthetic Data},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ABMIHYAW}},
  note         = {Machine review of arXiv:2507.09948}
}
abstract

Deep learning-based prediction models for High-Level Synthesis (HLS) of hardware designs often struggle to generalize. In this paper, we study how to close the generalizability gap of these models through pretraining on synthetic data and introduce Iceberg, a synthetic data augmentation approach that expands both large language model (LLM)-generated programs and weak labels of unseen design configurations. Our weak label generation method is integrated with an in-context model architecture, enabling meta-learning from actual and proximate labels. Iceberg improves the geometric mean modeling accuracy by $86.4\%$ when adapt to six real-world applications with few-shot examples and achieves a $2.47\times$ and a $1.12\times$ better offline DSE performance when adapting to two different test datasets. Our open-sourced code is here: https://github.com/UCLA-VAST/iceberg

Figures

Figures reproduced from arXiv: 2507.09948 by the authors.

Figure 1
Figure 1. Our approach to synthetic data augmentation spans both the program (P) dimension and the design configuration (T) dimension. To expand in the P dimension, we use LLMs to generate diverse synthetic programs, then obtain an actual labeled dataset by running DSE on these programs. To scale in the T dimension, we generate weak labels for unseen T configurations. Finally, we combine actual and weak labels to form a hybri… view at source ↗
Figure 2
Figure 2. presents our LLM-based framework for synthetic program generation. At its core is a dedicated initial prompt that defines all constraints and preferences for the desired program. To ensure diversity, multiple programs are iteratively generated for each ap￾plication domain. After generating a batch of programs, external tools are invoked to verify their validity. Common mistakes are then summarized, and the initial p… view at source ↗
Figure 3
Figure 3. Creating weak labels. Given the actual training data and the design space of each program in the training set, we train an ensemble of GNNs and apply random dropout to each ensemble member to produce a set of synthetic functions, denoted as {θj}j∈[n] . We then randomly sample a training program P, a GNN parameterized by θj , several unseen design configurations from the design space of P, and use the GNN to spawn we… view at source ↗
Figures from the paper (6 more)
Figure 4
Figure 4. Figure 4: Pretraining the G-TNP model. From the hybrid dataset containing both actual and weak labels, we sample sequences {(xi, yi)} and corre￾spond to the same program Pi and the same function θji . This ensure that the N points in each sequence share the same context. Then, w…
Figure 6
Figure 6. Figure 6: presents the results of our model-based forward optimization experiment, where we evaluate performance using the best@1 metric. (a) Adapting to HLSyn (b) Adapting to real world programs [PITH_FULL_IMAGE:figures/full_fig_p006_6.png]
Figure 8
Figure 8. Figure 8: Model accuracy rises when generating weak labels with a more diverse set of synthetic functions. The GNNs are trained with different randomly initialized parameters, which has been found to effectively produce a calibrated set of models [41]. In [PITH_FULL_IMAGE:figur…
Figure 7
Figure 7. Figure 7: compares the testing loss curve when using either a subset of HLSyn or Iceberg as the pretraining dataset and a subset of HLSyn as the testing dataset, with and without weak label augmentation. We observe a significant accuracy boost when weak labels are added, particu…
Figure 9
Figure 9. Figure 9: Comparison of Iceberg’s weak label generation approach with previous methods. Iceberg uses an ensemble of trained GNNs, while GP represents a randomly initialized Gaussian Process. Iceberg’s approach achieves significantly better results. The GP baseline is early stopp…
Figure 10
Figure 10. Figure 10: Different combination of HLS pragmas means different [PITH_FULL_IMAGE:figures/full_fig_p009_10.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

43 extracted references · 27 canonical work pages

  1. [1]

    High-level synthesis for FPGAs: From prototyping to deployment,

    J. Cong, B. Liu, S. Neuendorffer, J. Noguera, K. Vissers, and Z. Zhang, “High-level synthesis for FPGAs: From prototyping to deployment,” IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems, vol. 30, no. 4, pp. 473–491, 2011

    2011

  2. [2]

    FPGA HLS today: successes, challenges, and opportunities,

    J. Cong, J. Lau, G. Liu, S. Neuendorffer, P. Pan, K. Vissers, and Z. Zhang, “FPGA HLS today: successes, challenges, and opportunities,” ACM Transactions on Reconfigurable Technology and Systems (TRETS), vol. 15, no. 4, pp. 1–42, 2022

    2022

  3. [3]

    VerilogEval: Evaluating large language models for Verilog code generation,

    M. Liu, N. Pinckney, B. Khailany, and H. Ren, “VerilogEval: Evaluating large language models for Verilog code generation,” 2023. [Online]. Available: https://arxiv.org/abs/2309.07544

    arXiv 2023

  4. [4]

    MG-Verilog: Multi- grained dataset towards enhanced LLM-assisted Verilog generation,

    Y . Zhang, Z. Yu, Y . Fu, C. Wan, and Y . C. Lin, “MG-Verilog: Multi- grained dataset towards enhanced LLM-assisted Verilog generation,” in 2024 IEEE LLM Aided Design Workshop (LAD) . IEEE, 2024, pp. 1–5

    work page 2024

  5. [5]

    Origen: Enhancing RTL code generation with code-to-code augmentation and self-reflection,

    F. Cui, C. Yin, K. Zhou, Y . Xiao, G. Sun, Q. Xu, Q. Guo, D. Song, D. Lin, X. Zhang et al. , “Origen: Enhancing RTL code generation with code-to-code augmentation and self-reflection,” arXiv preprint arXiv:2407.16237, 2024

    arXiv 2024

  6. [6]

    ScaleHLS: A new scalable high-level synthesis framework on multi-level intermediate representation,

    H. Ye, C. Hao, J. Cheng, H. Jeong, J. Huang, S. Neuendorffer, and D. Chen, “ScaleHLS: A new scalable high-level synthesis framework on multi-level intermediate representation,” in 2022 IEEE Iternational Sym- posium on High-Performance Computer Architecture (HPCA) . IEEE, 2022, pp. 741–755

    work page 2022

  7. [7]

    Stream-HLS: Towards automatic dataflow acceleration,

    S. Basalama and J. Cong, “Stream-HLS: Towards automatic dataflow acceleration,” arXiv e-prints, pp. arXiv–2501, 2025

    work page 2025

  8. [8]

    LightningSimV2: Faster and scalable simulation for high-level synthesis via graph compilation and optimization,

    R. Sarkar, R. Paul, and C. C. Hao, “LightningSimV2: Faster and scalable simulation for high-level synthesis via graph compilation and optimization,” in 2024 IEEE 32nd Annual International Symposium on Field-Programmable Custom Computing Machines (FCCM) . IEEE, 2024, pp. 104–114

    work page 2024

Show all 43 references
  1. [9]

    A unified framework for automated code transformation and pragma insertion,

    S. Pouget, L.-N. Pouchet, and J. Cong, “A unified framework for automated code transformation and pragma insertion,” in Proceedings of the 2025 ACM/SIGDA International Symposium on Field Programmable Gate Arrays, 2025, pp. 187–198

    2025

  2. [10]

    Allo: A programming model for composable accelerator design,

    H. Chen, N. Zhang, S. Xiang, Z. Zeng, M. Dai, and Z. Zhang, “Allo: A programming model for composable accelerator design,” Proceedings of the ACM on Programming Languages , vol. 8, no. PLDI, pp. 593–620, 2024

    2024

  3. [11]

    An iteratively-refined dataset for high-level synthesis functional verification through LLM- aided bug injection,

    L. J. Wan, H. Ye, J. Wang, M. Jha, and D. Chen, “An iteratively-refined dataset for high-level synthesis functional verification through LLM- aided bug injection,” in 2024 IEEE LLM Aided Design Workshop (LAD). IEEE, 2024, pp. 1–6

    2024

  4. [12]

    HLSPilot: LLM-based high-level synthesis,

    C. Xiong, C. Liu, H. Li, and X. Li, “HLSPilot: LLM-based high-level synthesis,” 2024. [Online]. Available: https://arxiv.org/abs/2408.06810

    2024 arXiv

  5. [13]

    C2HLSC: Can LLMs bridge the software-to-hardware design gap?

    L. Collini, S. Garg, and R. Karri, “C2HLSC: Can LLMs bridge the software-to-hardware design gap?” arXiv preprint arXiv:2406.09233 , 2024

    2024 arXiv

  6. [14]

    Optimizing high-level synthesis designs with retrieval-augmented large language models,

    H. Xu, H. Hu, and S. Huang, “Optimizing high-level synthesis designs with retrieval-augmented large language models,” in 2024 IEEE LLM Aided Design Workshop (LAD) . IEEE, 2024, pp. 1–5

    2024

  7. [15]

    Automated C/C++ Program Repair for High-Level synthesis via Large Language Models,

    K. Xu, G. L. Zhang, X. Yin, C. Zhuo, U. Schlichtmann, and B. Li, “Automated C/C++ Program Repair for High-Level synthesis via Large Language Models,” in Proceedings of the 2024 ACM/IEEE International Symposium on Machine Learning for CAD , 2024, pp. 1–9

    2024

  8. [16]

    LLM-aided compilation for tensor accelerators,

    C. Hong, S. Bhatia, A. Haan, S. K. Dong, D. Nikiforov, A. Cheung, and Y . S. Shao, “LLM-aided compilation for tensor accelerators,” in 2024 IEEE LLM Aided Design Workshop (LAD) . IEEE, 2024, pp. 1–14

    2024

  9. [17]

    GPT4AIGchip: Towards next-generation AI accelerator design automation via large language models,

    Y . Fu, Y . Zhang, Z. Yu, S. Li, Z. Ye, C. Li, C. Wan, and Y . C. Lin, “GPT4AIGchip: Towards next-generation AI accelerator design automation via large language models,” in2023 IEEE/ACM International Conference on Computer Aided Design (ICCAD). IEEE, 2023, pp. 1–9

    2023

  10. [18]

    IRONMAN-PRO: Multiobjective design space exploration in HLS via reinforcement learning and graph neural network-based modeling,

    N. Wu, Y . Xie, and C. Hao, “IRONMAN-PRO: Multiobjective design space exploration in HLS via reinforcement learning and graph neural network-based modeling,” IEEE Transactions on Computer-Aided De- sign of Integrated Circuits and Systems , vol. 42, no. 3, pp. 900–913, 2022

    2022

  11. [19]

    Robust GNN-based representation learning for HLS,

    A. Sohrabizadeh, Y . Bai, Y . Sun, and J. Cong, “Robust GNN-based representation learning for HLS,” in 2023 IEEE/ACM International Conference on Computer Aided Design (ICCAD) , 2023, pp. 1–9

    2023

  12. [20]

    Balor: HLS source code evaluator based on custom graphs and hierarchical GNNs,

    E. Murphy and L. Josipovi ´c, “Balor: HLS source code evaluator based on custom graphs and hierarchical GNNs,” inIntl. Conference on Computer- Aided Design (ICCAD) , 2024, pp. 1–9

    2024

  13. [21]

    Hierarchical source-to-post-route QoR prediction in high-level synthesis with GNNs,

    M. Gao, J. Zhao, Z. Lin, and M. Guo, “Hierarchical source-to-post-route QoR prediction in high-level synthesis with GNNs,” in 2024 Design, Automation & Test in Europe Conference & Exhibition (DATE) . IEEE, 2024, pp. 1–6

    2024

  14. [22]

    Learning to compare hardware designs for high- level synthesis,

    Y . Bai, A. Sohrabizadeh, Z. Ding, R. Liang, W. Li, D. Wang, H. Ren, Y . Sun, and J. Cong, “Learning to compare hardware designs for high- level synthesis,” in Proceedings of the 2024 ACM/IEEE International Symposium on Machine Learning for CAD , 2024, pp. 1–7

    2024

  15. [23]

    PowerGear: Early-stage power estimation in FPGA HLS via heterogeneous edge- centric GNNs,

    Z. Lin, Z. Yuan, J. Zhao, W. Zhang, H. Wang, and Y . Tian, “PowerGear: Early-stage power estimation in FPGA HLS via heterogeneous edge- centric GNNs,” in 2022 Design, Automation & Test in Europe Confer- ence & Exhibition (DATE) . IEEE, 2022, pp. 1341–1346

    2022

  16. [24]

    Cross-modality program representation learning for electronic design automation with high-level synthesis,

    Z. Qin, Y . Bai, A. Sohrabizadeh, Z. Ding, Z. Hu, Y . Sun, and J. Cong, “Cross-modality program representation learning for electronic design automation with high-level synthesis,” in Proceedings of the 2024 ACM/IEEE International Symposium on Machine Learning for CAD , 2024, pp. 1–12

    2024

  17. [25]

    Hierarchical mixture of experts: Generalizable learning for high-level synthesis,

    W. Li, D. Wang, Z. Ding, A. Sohrabizadeh, Z. Qin, J. Cong, and Y . Sun, “Hierarchical mixture of experts: Generalizable learning for high-level synthesis,” in Proceedings of the AAAI Conference on Artificial Intelli- gence, vol. 39, no. 17, 2025, pp. 18 476–18 484

    2025

  18. [26]

    Efficient task transfer for HLS DSE,

    Z. Ding, A. Sohrabizadeh, W. Li, Z. Qin, Y . Sun, and J. Cong, “Efficient task transfer for HLS DSE,” in Proceedings of the 43rd IEEE/ACM International Conference on Computer-Aided Design , 2024, pp. 1–9

    2024

  19. [27]

    Towards a comprehensive benchmark for high-level synthesis targeted to FPGAs,

    Y . Bai, A. Sohrabizadeh, Z. Qin, Z. Hu, Y . Sun, and J. Cong, “Towards a comprehensive benchmark for high-level synthesis targeted to FPGAs,” Advances in Neural Information Processing Systems, vol. 36, pp. 45 288– 45 299, 2023

    2023

  20. [28]

    HLSFactory: A framework empowering high-level synthesis datasets for machine learning and beyond,

    S. Abi-Karam, R. Sarkar, A. Seigler, S. Lowe, Z. Wei, H. Chen, N. Rao, L. John, A. Arora, and C. Hao, “HLSFactory: A framework empowering high-level synthesis datasets for machine learning and beyond,” in Pro- ceedings of the 2024 ACM/IEEE International Symposium on Machine Le...

    2024

  21. [29]

    Source-to-source optimization for HLS,

    J. Cong, M. Huang, P. Pan, Y . Wang, and P. Zhang, “Source-to-source optimization for HLS,” FPGAs for Software Programmers, pp. 137–163, 2016

    2016

  22. [30]

    HIDA: A hierarchical dataflow compiler for high-level synthesis,

    H. Ye, H. Jun, and D. Chen, “HIDA: A hierarchical dataflow compiler for high-level synthesis,” in Proceedings of the 29th ACM International Conference on Architectural Support for Programming Languages and Operating Systems, Volume 1 , 2024, pp. 215–230

    2024

  23. [31]

    AutoDSE: Enabling software programmers to design efficient FPGA accelerators,

    A. Sohrabizadeh, C. H. Yu, M. Gao, and J. Cong, “AutoDSE: Enabling software programmers to design efficient FPGA accelerators,” ACM Trans. Des. Autom. Electron. Syst. , vol. 27, no. 4, Feb. 2022. [Online]. Available: https://doi.org/10.1145/3494534

    2022 doi

  24. [32]

    Transformer Neural Processes: Uncertainty- aware meta learning via sequence modeling,

    T. Nguyen and A. Grover, “Transformer Neural Processes: Uncertainty- aware meta learning via sequence modeling,” arXiv preprint arXiv:2207.04179, 2022

    2022 arXiv

  25. [33]

    Expt: Synthetic pretraining for few-shot experimental design,

    T. Nguyen, S. Agrawal, and A. Grover, “Expt: Synthetic pretraining for few-shot experimental design,” Advances in Neural Information Processing Systems, vol. 36, pp. 45 856–45 869, 2023

    2023

  26. [34]

    TabPFN: A transformer that solves small tabular classification problems in a second,

    N. Hollmann, S. M ¨uller, K. Eggensperger, and F. Hutter, “TabPFN: A transformer that solves small tabular classification problems in a second,” 2023. [Online]. Available: https://arxiv.org/abs/2207.01848

    2023 arXiv

  27. [35]

    Rosetta: A realistic high- level synthesis benchmark suite for software programmable FPGAs,

    Y . Zhou, U. Gupta, S. Dai, R. Zhao, N. Srivastava, H. Jin, J. Featherston, Y .-H. Lai, G. Liu, G. A. Velasquez et al. , “Rosetta: A realistic high- level synthesis benchmark suite for software programmable FPGAs,” in Proceedings of the 2018 ACM/SIGDA International Symposium o...

    2018

  28. [36]

    Attention is all you need,

    A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, Ł. Kaiser, and I. Polosukhin, “Attention is all you need,” Advances in Neural Information Processing Systems , vol. 30, 2017

    2017

  29. [37]

    Vitis HLS 2023.2,

    AMD/Xilinx, “Vitis HLS 2023.2,” https://docs.amd.com/r/en-US/ ug1399-vitis-hls/Introduction, 2023

    2023

  30. [38]

    Chain-of-Thought prompting elicits reasoning in large language models,

    J. Wei, X. Wang, D. Schuurmans, M. Bosma, B. Ichter, F. Xia, E. Chi, Q. Le, and D. Zhou, “Chain-of-Thought prompting elicits reasoning in large language models,” 2023. [Online]. Available: https://arxiv.org/abs/2201.11903

    2023 arXiv

  31. [39]

    LICO: Large language models for in-context molecular optimization,

    T. Nguyen and A. Grover, “LICO: Large language models for in-context molecular optimization,” arXiv preprint arXiv:2406.18851 , 2024

    2024

  32. [40]

    Neural processes,

    M. Garnelo, J. Schwarz, D. Rosenbaum, F. Viola, D. J. Rezende, S. Eslami, and Y . W. Teh, “Neural processes,” arXiv preprint arXiv:1807.01622, 2018

    2018 arXiv

  33. [41]

    Simple and scalable predictive uncertainty estimation using deep ensembles,

    B. Lakshminarayanan, A. Pritzel, and C. Blundell, “Simple and scalable predictive uncertainty estimation using deep ensembles,” Advances in neural information processing systems , vol. 30, 2017

    2017

  34. [42]

    Dropout as a Bayesian approximation: Representing model uncertainty in deep learning,

    Y . Gal and Z. Ghahramani, “Dropout as a Bayesian approximation: Representing model uncertainty in deep learning,” in International Conference on Machine Learning . PMLR, 2016, pp. 1050–1059

    2016

  35. [43]

    Mixture of in-context prompters for tabular PFNs,

    D. Xu, O. Cirit, R. Asadi, Y . Sun, and W. Wang, “Mixture of in-context prompters for tabular PFNs,” arXiv preprint arXiv:2405.16156 , 2024. APPENDIX A. Detailed experiment results Table VI presents the performance of HARP and H-MoE on HLSyn, with and without pretraining on Ic...

    2024 arXiv

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.