Pith. sign in

REVIEW 3 major objections 6 minor 26 references

CORE: Constraint-Aware One-Step Reinforcement Learning for Simulation-Guided Neural Network Accelerator Design

T0 review · 3 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read CORE claims that a critic-free one-step RL policy can outperform a genetic algorithm and HASCO in DNN accelerator hardware-mapping co-design, achieving at least a 15x improvement in optimal reward within a 40,000-sample budget.

desk verdict Solid one-step RL extension with a useful constrained decoder, but the 15x claim rests on a handicapped HASCO baseline and single-run numbers. read the letter →

arxiv 2506.03474 v1 pith:YVOFLC27 submitted 2025-06-04 cs.LG cs.AIcs.AR

Yifeng Xiao , Yurong Xu , Ning Yan , Masood Mortazavi , Pierluigi Nuzzo This is my paper
classification cs.LGcs.AIcs.AR
keywords one-stepreinforcementlearningdesignspaceexplorationDNNacceleratorco-designhardware-mappingscaling-graphdecoderrewardshapingsampleefficiencyconstraint-awareoptimization
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

The paper tries to establish that simulation-based design-space exploration can be made dramatically more sample-efficient by treating each candidate design as a one-step action of a policy, rather than as a rollout of many sequential decisions. It proposes CORE, a critic-free RL method: the policy samples a complete accelerator configuration from a factorized distribution, a scaling-graph decoder enforces parameter dependencies, and reward shaping penalizes invalid or anomalous designs. Applied to hardware-mapping co-design of DNN accelerators, CORE is claimed to find better latency and latency-area configurations than a genetic algorithm and HASCO with a fixed 40,000-sample budget, achieving at least a 15x improvement in the optimal reward. The sympathetic reader should care because the recipe is not tied to accelerators: any expensive, constrained, black-box design space could in principle be searched the same way.

What carries the argument

The machinery is a one-step MDP: instead of making design choices sequentially, the agent draws the complete compound action in a single step from a policy that factorizes into conditional probability density functions, using categorical distributions for small discrete choices and Beta distributions for large or context-dependent ranges. The scaling-graph decoder then converts the sampled actions into a feasible configuration: it is a directed graph whose edges say that the decoded value of a source parameter caps the feasible range of a dependent target parameter, and decoding follows topological order so constraints like tile-size hierarchy and PE-bounded parallelism hold. Reward shaping handles the simulator feedback: constraint violations subtract a scaled penalty, and designs that cannot be simulated get a reward pushed below the running average. The policy update uses a surrogate objective with a batch-relative advantage, a KL regularizer to keep updates close to the current policy, and a decaying entropy bonus.

What would settle it

Take the CORE-optimized configuration for ResNet-50 on the cloud platform and evaluate it, together with the genetic-algorithm and HASCO configurations, under a cycle-accurate RTL simulator or a fabricated chip, comparing measured latency and area against the MAESTRO and synthesis predictions; if the real ranking reverses or the 15x gap largely disappears, the central claim fails. A cheaper probe is to rerun CORE with a different analytical cost model, such as Timeloop, and check whether CORE still dominates the baselines.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central discovery is that the one-step RL formulation removes the two bottlenecks of multi-step RL for design search: sparse delayed rewards and the need for a value function. By sampling a whole design at once, evaluating a parallel batch, and defining each sample's advantage as its reward minus a running average of batch rewards, CORE turns expensive simulation feedback into immediate policy-gradient signal without critics. The scaling-graph decoder makes this feasible in practice: it maps sampled continuous actions to valid discrete parameters by capping each dependent parameter's range with already-decoded source values, so most samples respect structural constraints. The paper reports that CORE consistently beats GA and HASCO on larger DNNs and more complex design spaces, with at least a 15x improvement in optimal reward, and that ablations removing either reward shaping or scaling-graph decoding degrade performance, in some edge cases leaving no valid design found at all.

Load-bearing premise

The whole result rests on the simulated cost models (MAESTRO for latency and the synthesis flow for area) being an accurate proxy for real accelerator performance; the paper's own Discussion concedes that its effectiveness depends on the fidelity of these cost models.

Editorial extensions

If this is right

  • CORE's one-step, batch-relative policy update removes value-function learning, so training cost scales with the number of simulations rather than rollout length, and the same 40,000-sample budget found better designs than the baselines.
  • CORE requires fewer samples to reach a given latency or latency-area value than the genetic algorithm on the reported curves, so design-space exploration cost drops when simulation is the bottleneck.
  • Reward shaping converts invalid or anomalous designs into quantitative penalties, so the agent receives feedback on every sample instead of discarding failures, improving exploration under constraints.
  • The combination of a scaling-graph decoder and reward shaping is the source of the gain in the ablations: removing either one degrades results, and removing reward shaping can prevent any valid design from being found on edge platforms.
  • Because the method is formulated independently of the accelerator simulator, the same code path applies to any structured design problem expressible as a one-step MDP with constraints.

Reading between the lines

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

  • If the cost models are faithful, the same one-step update should transfer to other expensive constrained search spaces; a direct test would be applying CORE to compiler flag tuning or robot co-design and checking whether its sample-efficiency edge survives outside accelerators, since the paper only demonstrates the accelerator instance.
  • The batch-relative advantage is essentially a within-batch comparison, which suggests a connection to preference-based optimization where rewards are replaced by pairwise rankings; a testable variant would feed CORE only ordinal comparisons of designs and see whether the 15x gain persists without absolute rewards.
  • The static, compile-time mapping assumption means the result does not yet cover runtime-adaptive dataflows; extending the scaling graph to dynamic decisions is a natural stress test the paper explicitly leaves for future work.
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

3 major / 6 minor

Summary. The paper proposes CORE, a critic-free one-step reinforcement learning method for simulation-guided design space exploration, instantiated for DNN accelerator hardware-mapping co-design. The policy factorizes a joint distribution over design actions using conditional Beta and categorical distributions, decodes actions into feasible configurations via a scaling-graph-based decoder, and is updated with a surrogate objective based on batch-relative rewards without a value function. Experiments compare CORE with GA and HASCO on seven DNN models under cloud and edge area constraints, together with ablations of the reward shaping and scaling-graph decoding, and the paper claims 'at least a 15× improvement in the optimal reward' and better sample efficiency.

Significance. If the empirical comparison is made sound, the paper would be a useful contribution: the one-step, critic-free formulation is well matched to expensive, constrained design-space exploration, and the scaling-graph decoder plus constraint-aware reward shaping are sensible and clearly described. The problem statement, algorithm description, and ablation design are coherent, and the Discussion honestly acknowledges the dependence on cost-model fidelity and the relation to concurrent work on batch-relative advantage. The main obstacles are in the experimental validation rather than in the algorithmic derivation.

major comments (3)
  1. [Section 4.2, Table 1] The central claim of 'at least a 15× improvement in the optimal reward' is not supported by the comparison as configured. The paper states that 'Because the space defined in Figure 1c is too large for HASCO, we limit it by using larger step sizes when discretizing the ranges of the number of PEs and the buffer sizes.' This means HASCO is not searching the same design space at the same granularity as CORE; the log10 gaps in Table 1 therefore conflate algorithmic quality with search-space resolution. To support the headline claim, HASCO (and GA) should be run on CORE's full-resolution space, or the claim should be restricted to the coarse-baseline comparison.
  2. [Section 4.2, Table 1 and Figure 6] All results appear to be single runs: no seeds, standard deviations, or confidence intervals are reported anywhere in Table 1 or Figure 6. Since CORE samples from stochastic Beta/categorical distributions and GA is randomized, a single trajectory cannot establish a factor-of-15 superiority or the sample-efficiency comparisons. Please report means and variance across multiple seeds (at least 5) and provide a reproducibility statement with the experimental settings.
  3. [Section 4.2, 'Main Results' paragraph] The sentence 'While HASCO achieves better metrics on smaller DNNs like NCF and DLRM, CORE consistently obtains optimal results across other cases' is inconsistent with Table 1. In the latency columns CORE is better than HASCO on NCF and DLRM on both platforms (e.g., NCF cloud: 2.17 vs 3.48), while HASCO is better only in some LAS rows (e.g., NCF edge LAS: 3.65 vs 4.06). This discrepancy needs to be corrected, and the claim 'consistently obtains optimal results' should be qualified per objective and per model.
minor comments (6)
  1. [Section 4.1, hyperparameters] No hyperparameter sensitivity is reported for the reward-shaping and regularization coefficients (α_c, α_p, α_r, β_e, β_r) despite the statement 'No hyperparameter tuning is performed'; a small sensitivity study would make the ablation conclusions more robust.
  2. [Section 4.2] The definition of the 15× improvement should be explicit: state whether it is a ratio of raw metrics, of geometric means of the log10-scaled values, or something else.
  3. [Table 1] The caption and text should state unambiguously that lower log10 values are better for both latency and LAS; the narrative around 'better metrics' is currently easy to misread.
  4. [Figure 6] HASCO is excluded from the sample-efficiency comparison with the justification that it is two-stage; this is reasonable, but the text should tie this exclusion back to the 15× claim so readers do not infer that HASCO was beaten on sample efficiency.
  5. [Broader Impacts] The text states 'open-access implementation' but no code repository or link is provided; please add one or remove the statement.
  6. [Section 3.2, Eq. (4)] The scaling-graph decoding is described for upper-bound constraints; the paper does not specify how lower-bound constraints or non-monotone dependencies are handled, though these appear in the motivating example (e.g., P1 ≥ P_low).

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: CORE's critic-free one-step formulation is independently corroborated (Ghraieb et al. [7]; DeepSeek-R1 [10], which the paper itself acknowledges), and its central claims are evaluated against external baselines and external cost models, so the 15x claim is not forced by construction.

full rationale

The derivation chain is self-contained and empirically anchored. CORE's one-step policy (Eq. 2), scaling-graph decoding (Eqs. 3-5), constraint-aware reward shaping (Eqs. 10-11), and batch-relative surrogate advantage (Eqs. 7-9, 12-14) introduce no fitted parameter that is later renamed a prediction: the policy is trained on the raw simulator reward w^T U(xi), and the Table 1 headline numbers are raw log10 latency and log10 latency-area-sum of the best sampled design, not the shaped training reward, so the optimization target is not defined in terms of the method's own mechanism. The only self-citations ([8,9], prior single-step RL work by two of the present authors) supply the KL-regularized, entropy-decayed surrogate-objective design in Section 3.3, but the same core idea (advantage computed by intra-batch reward comparison without a critic) is independently and externally stated in DeepSeek-R1 [10], which the paper itself flags in the Discussion ('a recent method [10] independently proposes a similar idea to compute the advantage via intra-batch reward comparisons'), so the self-citation is corroborated rather than load-bearing. The central empirical claims are tested against external baselines (GA, HASCO) under an external evaluation engine (MAESTRO cost model plus Synopsys DC and Cadence Innovus area estimation), so no result reduces to its own inputs by construction. The remaining concerns are benchmark-fairness and external-validity issues rather than circularity: HASCO was deliberately restricted to coarser discretization ('Because the space defined in Figure 1c is too large for HASCO, we limit it by using larger step sizes'), results appear to be single runs without variance reporting, and the cost-model fidelity limitation is acknowledged in the Discussion ('its effectiveness depends on the fidelity of the underlying cost models'). None of these is a definitional or self-citational reduction of the claimed improvement. The score of 1 reflects only the presence of minor, non-load-bearing self-citation in the surrogate-objective design.

Assumptions & free parameters 7 free parameters · 3 assumptions · 0 invented entities

The central claim rests on user-chosen reward weights, penalty rates, and the assumption that the MAESTRO simulator is an accurate proxy for real hardware. No free parameters are fitted to the test results; they are fixed before training. The only invented structure is the scaling graph, which is a decoding mechanism, not a postulated physical entity.

free parameters (7)
  • Reward weight vector w = not reported
    Eq. 6 defines R as w^T U; the user-defined weights for latency, area, and power set the optimization trade-off and are chosen by hand (Section 3.3).
  • Violation penalty rate alpha_c = 1.0
    Section 4.1 states 'other rate factors are fixed at 1.0'; it scales the penalty in Eq. 10 for constraint violations.
  • Anomalous design penalty rate alpha_p = 1.0
    Scales the penalty for nonsimulatable designs in Eq. 11; fixed at 1.0.
  • Renewal rate alpha_r = 0.2
    Controls the exponential moving average of batch rewards in Eq. 7; reported in Section 4.1.
  • Entropy coefficient beta_e = decayed from 1.0 to 0.02
    Entropy regularization weight in Eq. 14; linearly decayed schedule set by hand.
  • KL factor beta_r = 1.0
    KL regularization weight in Eq. 13; 'other rate factors are fixed at 1.0' (Section 4.1).
  • Batch size and learning rate = E=32, lr=1e-5
    Reported in Section 4.1; number of parallel simulations per episode and Adam step size.
assumptions (3)
  • domain assumption The MAESTRO cost model and the Synopsys DC/Cadence Innovus flow provide metrics that are faithful to real accelerator performance.
    Section 4.1 uses MAESTRO for latency and area; the Discussion states effectiveness depends on cost-model fidelity.
  • domain assumption All design dependencies can be represented as upper-bound scaling relations in a directed graph.
    Section 3.2 Eq. 4 replaces the upper bound with min(A_i); constraints that are not simple upper bounds would not be enforced.
  • domain assumption The design problem is a one-step, stateless MDP in which a full design can be sampled in one action.
    Definition 1 fixes a single state s0; this excludes sequential or adaptive design decisions.

how reviews work

0 comments
Cite this review

Pith. "Pith review of CORE: Constraint-Aware One-Step Reinforcement Learning for Simulation-Guided Neural Network Accelerator Design." pith.science (2026). https://pith.science/paper/YVOFLC27

@misc{pith2026250603474,
  author       = {Pith},
  title        = {Pith review of: CORE: Constraint-Aware One-Step Reinforcement Learning for Simulation-Guided Neural Network Accelerator Design},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/YVOFLC27}},
  note         = {Machine review of arXiv:2506.03474}
}
read the original abstract

Simulation-based design space exploration (DSE) aims to efficiently optimize high-dimensional structured designs under complex constraints and expensive evaluation costs. Existing approaches, including heuristic and multi-step reinforcement learning (RL) methods, struggle to balance sampling efficiency and constraint satisfaction due to sparse, delayed feedback, and large hybrid action spaces. In this paper, we introduce CORE, a constraint-aware, one-step RL method for simulationguided DSE. In CORE, the policy agent learns to sample design configurations by defining a structured distribution over them, incorporating dependencies via a scaling-graph-based decoder, and by reward shaping to penalize invalid designs based on the feedback obtained from simulation. CORE updates the policy using a surrogate objective that compares the rewards of designs within a sampled batch, without learning a value function. This critic-free formulation enables efficient learning by encouraging the selection of higher-reward designs. We instantiate CORE for hardware-mapping co-design of neural network accelerators, demonstrating that it significantly improves sample efficiency and achieves better accelerator configurations compared to state-of-the-art baselines. Our approach is general and applicable to a broad class of discrete-continuous constrained design problems.

Figures

Figures reproduced from arXiv: 2506.03474 by the authors.

Figure 1
Figure 1. (a) Hardware resources for a 2-level spatial DNN accelerator; (b) Tensor dimensions for [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Overview of the CORE framework for simulation-based design space exploration. function R is defined as R(s0, a) : {s0} × A → R, a ∈ A. The agent interacts with the environment in episodes of length 1. We model the design exploration process as a one-step MDP, where a policy network πθ represents a joint distribution over design actions conditioned on a fixed input state s0. Since the environment is stateless, s0 act… view at source ↗
Figure 3
Figure 3. Evaluation pipeline for hardware-mapping co-design. [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Scaling graph for DNN accelerator. Scaling-Graph-Based Decoding. To effectively navigate structured parameter spaces with interde￾pendencies, we introduce a decoding strategy based on a scaling graphs, which dynamically adjusts the parameter bounds, improving feasibili…
Figure 5
Figure 5. Figure 5: Average target metrics on different platforms for different methods. [PITH_FULL_IMAGE:figures/full_fig_p009_5.png]
Figure 6
Figure 6. Figure 6: Optimized latency and LAS as the number of samples [PITH_FULL_IMAGE:figures/full_fig_p009_6.png]
Figure 7
Figure 7. Figure 7: Notation for the dimensions used in convolution operations. [PITH_FULL_IMAGE:figures/full_fig_p012_7.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

26 extracted references · 20 canonical work pages

  1. [1]

    Genetic algorithms.Scientific american, 267(1):66–73, 1992

    John H Holland. Genetic algorithms.Scientific american, 267(1):66–73, 1992

    1992

  2. [2]

    Bayesian optimization with adaptive surrogate models for automated experimental design.Npj Computational Materials, 7(1):194, 2021

    Bowen Lei, Tanner Quinn Kirk, Anirban Bhattacharya, Debdeep Pati, Xiaoning Qian, Raymundo Arroyave, and Bani K Mallick. Bayesian optimization with adaptive surrogate models for automated experimental design.Npj Computational Materials, 7(1):194, 2021

    work page 2021

  3. [3]

    Reinforcement learning: A survey

    Leslie Pack Kaelbling, Michael L Littman, and Andrew W Moore. Reinforcement learning: A survey. Journal of artificial intelligence research, 4:237–285, 1996

    1996

  4. [4]

    Proximal policy optimization algorithms.arXiv preprint arXiv:1707.06347, 2017

    John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. Proximal policy optimization algorithms.arXiv preprint arXiv:1707.06347, 2017

    arXiv 2017

  5. [5]

    Archgym: An open-source gymnasium for machine learning assisted architecture design

    Srivatsan Krishnan, Amir Yazdanbakhsh, Shvetank Prakash, Jason Jabbour, Ikechukwu Uchendu, Susobhan Ghosh, Behzad Boroujerdian, Daniel Richins, Devashree Tripathy, Aleksandra Faust, et al. Archgym: An open-source gymnasium for machine learning assisted architecture design. InProceedings of the 50th Annual International Symposium on Computer Architecture, ...

    work page 2023

  6. [6]

    Delving into macro placement with reinforcement learning

    Zixuan Jiang, Ebrahim Songhori, Shen Wang, Anna Goldie, Azalia Mirhoseini, Joe Jiang, Young-Joon Lee, and David Z Pan. Delving into macro placement with reinforcement learning. In2021 ACM/IEEE 3rd Workshop on Machine Learning for CAD (MLCAD), pages 1–3. IEEE, 2021

    work page 2021

  7. [7]

    Single-step deep reinforcement learning for open-loop control of laminar and turbulent flows.Physical Review Fluids, 6(5):053902, 2021

    Hassan Ghraieb, Jonathan Viquerat, Aurélien Larcher, Philippe Meliga, and Elie Hachem. Single-step deep reinforcement learning for open-loop control of laminar and turbulent flows.Physical Review Fluids, 6(5):053902, 2021

    work page 2021

  8. [8]

    Mortazavi, Tiancheng Qin, and Ning Yan

    Masood S. Mortazavi, Tiancheng Qin, and Ning Yan. Theta-resonance: A single-step reinforcement learning method for design space exploration, 2022

    work page 2022

Show all 26 references
  1. [9]

    Mortazavi, and Ning Yan

    Jiayu Li, Masood S. Mortazavi, and Ning Yan. Exploring distributed circuit design using single-step reinforcement learning. In2024 60th ACM/IEEE Design Automation Conference (DAC). IEEE, 2024

    2024

  2. [10]

    Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning.arXiv preprint arXiv:2501.12948, 2025

    Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Ruoyu Zhang, Runxin Xu, Qihao Zhu, Shirong Ma, Peiyi Wang, Xiao Bi, et al. Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning.arXiv preprint arXiv:2501.12948, 2025

    2025 arXiv

  3. [11]

    Eyeriss: An energy-efficient reconfigurable accelerator for deep convolutional neural networks.IEEE journal of solid-state circuits, 52(1):127–138, 2016

    Yu-Hsin Chen, Tushar Krishna, Joel S Emer, and Vivienne Sze. Eyeriss: An energy-efficient reconfigurable accelerator for deep convolutional neural networks.IEEE journal of solid-state circuits, 52(1):127–138, 2016

    2016

  4. [12]

    Nvdla deep learning accelerator.http://nvdla.org, 2017

    NVDLA Project. Nvdla deep learning accelerator.http://nvdla.org, 2017

    2017

  5. [13]

    Shidiannao: Shifting vision processing closer to the sensor

    Zidong Du, Robert Fasthuber, Tianshi Chen, Paolo Ienne, Ling Li, Tao Luo, Xiaobing Feng, Yunji Chen, and Olivier Temam. Shidiannao: Shifting vision processing closer to the sensor. InProceedings of the 42nd annual international symposium on computer architecture, pages 92–104, 2015

    2015

  6. [14]

    Simulated annealing.Statistical science, 8(1):10–15, 1993

    Dimitris Bertsimas and John Tsitsiklis. Simulated annealing.Statistical science, 8(1):10–15, 1993

    1993

  7. [15]

    Towards automated risc-v mi- croarchitecture design with reinforcement learning

    Chen Bai, Jianwang Zhai, Yuzhe Ma, Bei Yu, and Martin DF Wong. Towards automated risc-v mi- croarchitecture design with reinforcement learning. InProceedings of the AAAI Conference on Artificial Intelligence, volume 38, pages 12–20, 2024

    2024

  8. [16]

    Confuciux: Autonomous hardware resource as- signment for dnn accelerators using reinforcement learning

    Sheng-Chun Kao, Geonhwa Jeong, and Tushar Krishna. Confuciux: Autonomous hardware resource as- signment for dnn accelerators using reinforcement learning. In2020 53rd Annual IEEE/ACM International Symposium on Microarchitecture (MICRO), pages 622–636. IEEE, 2020

    2020

  9. [17]

    Gamma: Automating the hw mapping of dnn models on accelerators via genetic algorithm

    Sheng-Chun Kao and Tushar Krishna. Gamma: Automating the hw mapping of dnn models on accelerators via genetic algorithm. InProceedings of the 39th International Conference on Computer-Aided Design, pages 1–9, 2020

    2020

  10. [18]

    Flextensor: An automatic schedule exploration and optimization framework for tensor computation on heterogeneous system

    Size Zheng, Yun Liang, Shuo Wang, Renze Chen, and Kaiwen Sheng. Flextensor: An automatic schedule exploration and optimization framework for tensor computation on heterogeneous system. InProceedings of the Twenty-Fifth International Conference on Architectural Support for Prog...

    2020

  11. [19]

    Digamma: Domain-aware genetic algorithm for hw-mapping co-optimization for dnn accelerators

    Sheng-Chun Kao, Michael Pellauer, Angshuman Parashar, and Tushar Krishna. Digamma: Domain-aware genetic algorithm for hw-mapping co-optimization for dnn accelerators. InDATE, 2022

    2022

  12. [20]

    Hasco: To- wards agile hardware and software co-design for tensor computation

    Qingcheng Xiao, Size Zheng, Bingzhe Wu, Pengcheng Xu, Xuehai Qian, and Yun Liang. Hasco: To- wards agile hardware and software co-design for tensor computation. In2021 ACM/IEEE 48th Annual International Symposium on Computer Architecture (ISCA), pages 1055–1068. IEEE, 2021

    2021

  13. [21]

    Unico: Unified hardware software co-optimization for robust neural network acceleration

    Bahador Rashidi, Chao Gao, Shan Lu, Zhisheng Wang, Chunhua Zhou, Di Niu, and Fengyu Sun. Unico: Unified hardware software co-optimization for robust neural network acceleration. InProceedings of the 56th Annual IEEE/ACM International Symposium on Microarchitecture, pages 77–90, 2023

    2023

  14. [22]

    A mechanistic performance model for superscalar out-of-order processors.ACM Transactions on Computer Systems (TOCS), 27(2):1–37, 2009

    Stijn Eyerman, Lieven Eeckhout, Tejas Karkhanis, and James E Smith. A mechanistic performance model for superscalar out-of-order processors.ACM Transactions on Computer Systems (TOCS), 27(2):1–37, 2009

    2009

  15. [23]

    Revisiting design choices in proximal policy optimization.arXiv preprint arXiv:2009.10897, 2020

    Chloe Ching-Yun Hsu, Celestine Mendler-Dünner, and Moritz Hardt. Revisiting design choices in proximal policy optimization.arXiv preprint arXiv:2009.10897, 2020

    2009 arXiv

  16. [24]

    Simba: Scaling deep-learning inference with multi-chip-module-based architecture

    Yakun Sophia Shao, Jason Clemons, Rangharajan Venkatesan, Brian Zimmer, Matthew Fojtik, Nan Jiang, Ben Keller, Alicia Klinefelter, Nathaniel Pinckney, Priyanka Raina, et al. Simba: Scaling deep-learning inference with multi-chip-module-based architecture. InProceedings of the ...

    2019

  17. [25]

    Maestro: A data-centric approach to understand reuse, performance, and hardware cost of dnn mappings.IEEE micro, 40(3):20–29, 2020

    Hyoukjun Kwon, Prasanth Chatarasi, Vivek Sarkar, Tushar Krishna, Michael Pellauer, and Angshuman Parashar. Maestro: A data-centric approach to understand reuse, performance, and hardware cost of dnn mappings.IEEE micro, 40(3):20–29, 2020

    2020

  18. [26]

    Timeloop: A systematic approach to dnn accelerator evaluation

    Angshuman Parashar, Priyanka Raina, Yakun Sophia Shao, Yu-Hsin Chen, Victor A Ying, Anurag Mukkara, Rangharajan Venkatesan, Brucek Khailany, Stephen W Keckler, and Joel Emer. Timeloop: A systematic approach to dnn accelerator evaluation. In2019 IEEE international symposium on ...

    2019

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.