Pith. sign in

REVIEW 3 major objections 5 minor 55 references

DOSA: Differentiable Model-Based One-Loop Search for DNN Accelerators

T0 review · 3 major / 5 minor · reviewed 2026-08-04 · deepseek-v4-flash

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

Pith's one-line read A single differentiable search loop that co-optimizes DNN accelerator hardware and mappings beats random and Bayesian search by 2.80x and 12.59x in energy-delay product.

desk verdict Good new technique, but the sample-efficiency headline overcounts by equating cheap gradient steps with full simulations. read the letter →

arxiv 2509.10702 v1 pith:SQZAYXRO submitted 2025-09-12 cs.AR cs.LG

Charles Hong , Qijing Huang , Grace Dinh , Mahesh Subedar , Yakun Sophia Shao This is my paper
classification cs.ARcs.LG
keywords designspaceexplorationDNNacceleratorhardware-softwareco-designdifferentiableperformancemodelmapping-firstsearchgradientdescentenergy-delayproductone-loop
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's thesis is that hardware-software co-design for DNN accelerators need not be a combinatorial two-loop search. By writing a closed-form differentiable model of latency and energy as functions of per-layer tiling factors, and by deriving the minimal hardware that supports a given mapping, the authors collapse the joint space into a single optimization landscape over mappings. Gradient descent over all layers simultaneously yields accelerator designs that, at a similar number of sampled evaluations, improve energy-delay product by 2.80x over random search and 12.59x over Bayesian optimization. The same machinery, augmented with a learned correction trained on RTL measurements, improves a real accelerator's EDP by 1.82x over its hand-tuned configuration. The sympathetic reader takes away that interpretable analytical models and gradient-based optimization can replace brute-force sampling in accelerator design space exploration.

What carries the argument

The load-bearing object is the differentiable analytical model over mapping variables. For each layer, spatial and temporal tiling factors (the loop iteration bounds at each memory level) determine (1) the minimal PE, accumulator, and scratchpad capacities via the parameter-wise max over layers, and (2) the write/update/read traffic at each memory level, from which latency and energy are assembled into the model-level EDP loss. Gradient descent with rounding to valid divisors searches all layers' tiling factors in parallel; loop ordering is handled separately by an iterative or softmax-weighted selection among weight-, input-, and output-stationary orders. The learned-correction variant augm

What would settle it

Run a head-to-head search with the same 10,000-sample budget on a workload whose best-known implementation fuses several layers to share buffers and bandwidth; if a pipeline-aware two-loop searcher finds an EDP more than 2.80x below DOSA's best, the parameter-wise-max minimal-hardware reduction—not gradient descent—is the bottleneck.

Watch

Extended reading notes

Core claim

The central claim is that the mapping-first reduction makes DSE differentiable: if hardware parameters are always set to the pointwise maximum of the minimal requirements implied by each layer's mapping, then the hardware search loop disappears and the objective becomes a smooth function of the mapping variables alone. DOSA constructs such a function from closed-form write/update/read traffic counts, roofline-style latency with a max over memory levels, and energy from per-access costs, then minimizes the full-model energy-delay product via gradient descent on all layers' spatial and temporal tiling factors. Evaluated on four diverse DNN workloads, this finds co-design points whose EDP is 2.

Load-bearing premise

The whole construction assumes the best accelerator for a multi-layer network is exactly the pointwise maximum of the minimal hardware each layer's mapping requires; the paper fixes weight-stationary dataflow, fixes tensor bypassing, and explicitly does not search pipeline-fused mappings, so any optimum that needs larger buffers for pipelining, a different dataflow, or shared-bandwidth scheduling is outside DOSA's search space.

Editorial extensions

If this is right

  • At roughly 10,000 model evaluations, DOSA's best co-design points achieve 2.80x lower EDP than random search and 12.59x lower than Bayesian optimization across the four evaluated workloads.
  • The hardware designs DOSA selects are themselves better: with a fixed strong mapper, they reduce EDP by 3.21x over the randomly seeded hardware start points.
  • On the hardware DOSA picks, its mappings beat the fixed mapper by 1.79x and a 1000-sample random mapper by 2.78x.
  • Augmenting the analytical model with a learned latency correction trained on 1,567 RTL samples gives a 1.82x EDP improvement over the target accelerator's default buffer sizes and mapping heuristic.
  • Gradient-based loop-ordering search gives 1.58x EDP improvement over no ordering search; iterative selection of the three stationary orderings gives 1.70x.

Reading between the lines

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

  • Because the parameter-wise-max reduction forces hardware to equal each layer's minimum, the reported gains do not cover designs where the optimum requires deliberate over-provisioning (e.g., for multi-layer pipelining, double buffering, or multi-tenancy); a pipeline-aware search could close or reverse the gap on latency-bound streaming workloads.
  • The same differentiable-model trick should transfer to other search objectives (area, throughput) and other dataflows; the fixed weight-stationary dataflow and fixed tensor bypassing are conveniences, not requirements of the mathematical formulation.
  • The 12.59x margin over Bayesian optimization likely reflects BO spending a large sample budget on hardware points before mapping optimization; embedding DOSA's differentiable model as an acquisition-function prior inside BO could be a strictly stronger baseline.
  • The learned-correction recipe suggests a practical path for accelerator vendors: keep the interpretable analytical model for optimization and periodically refit a small residual network from RTL measurements as the design converges, avoiding the need to train a full black-box surrogate.
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 / 5 minor

Summary. The paper introduces DOSA, a mapping-first, one-loop design-space exploration framework for DNN accelerators. DOSA converts the Timeloop/Accelergy analytical performance model for a Gemmini-like weight-stationary accelerator into closed-form differentiable equations for latency, energy, and EDP, then uses gradient descent (Adam) to co-optimize per-layer spatial/temporal tiling factors and the minimal hardware parameters (PE array size, buffer sizes) implied by those mappings. A DNN correction model is also trained on FireSim RTL latencies and used to augment the analytical model. The paper reports 0.18% MAE against Timeloop on 10,000 random mappings, claims 2.80x and 12.59x EDP improvements over random search and Bayesian optimization at similar sample counts, shows final Timeloop-evaluated designs beating Eyeriss/NVDLA/Gemmini baselines, and reports 1.82x EDP improvement over the default Gemmini-RTL configuration when the DNN-augmented model is used.

Significance. If the claims hold, this is a meaningful contribution: it is one of the first demonstrations that a differentiable analytical performance model can enable gradient-based, all-layer simultaneous optimization of mappings and hardware parameters, and it includes a reproducible artifact and final evaluation with Timeloop and FireSim rather than only the surrogate loss. The model accuracy, the explicit validation on a large random mapping set, and the RTL-backed case study are concrete strengths. The main risk is that the headline sample-efficiency comparison counts closed-form model evaluations and Timeloop simulations as equivalent, which is not a like-for-like comparison; this needs rework before the quantitative claims can be accepted.

major comments (3)
  1. [Section 6.3, Fig. 7] The headline '2.80x/12.59x with a similar number of samples' rests on the sentence 'Evaluations done using Timeloop are considered equivalent to evaluations done using DOSA's differentiable model.' Per Section 6.1, DOSA's curve uses 7 start points x 1490 Adam steps = ~10,430 closed-form model evaluations, each costing a few floating-point operations, while the random baseline uses 10 hardware designs x 1000 mappings = 10,000 Timeloop/Accelergy simulations, and BB-BO uses 100 hardware designs x 100 mappings per layer plus GP acquisition candidates. These are not commensurable samples. The claim of sample efficiency is therefore not established. Please either report wall-clock time, or compare methods under an equal budget of expensive (Timeloop) evaluations, or re-label the x-axis as 'model evaluations' and clearly separate the cost of a surrogate evaluation from the cost of a simulator e
  2. [Section 4.1, Eqs. (1)-(5); Section 5.1, Section 5.3.4] The hardware space is restricted to the parameter-wise max of each layer's minimal hardware requirements (Eq. 5 and Figure 3), with fixed weight-stationary C-K spatial dataflow, fixed bypassing, and no pipeline fusion. This is acknowledged in Sections 5.1 and 5.3.4, but the abstract and introduction present DOSA as simultaneously searching the hardware and mapping spaces in general. Under this restriction, hardware that is intentionally larger than the single-layer minimum (e.g., for pipelining, double buffering, multi-tenancy, or dataflow-specific buffering) is outside the search space. This is a scope limitation, not an internal inconsistency, but it should be stated explicitly in the title/abstract framing and ideally tested by ablating the minimal-parameterization assumption on at least one workload.
  3. [Section 4.6 vs. Sections 5.3.1 and 6.3] The differentiable model is validated on 10,000 random mappings (Fig. 4), but the optimizer's start-point rejection (Section 5.3.1) and gradient descent traverse a non-random subset of the mapping space, where model error could be higher. The final Timeloop/FireSim evaluations of the selected designs mitigate this concern for the final result, but the sample-efficiency comparison in Fig. 7 is still based on the surrogate's internal objective. Please report the differentiable model's error on the distribution of mappings actually visited by gradient descent and on the rejected start points, to confirm that the search is not exploiting model artifacts.
minor comments (5)
  1. [Abstract and Section 6.3] The text uses 'samples' and 'model evaluations' interchangeably. Figure 7's x-axis is 'model evaluations' while the abstract and Section 6.3 say 'samples.' Please harmonize terminology to avoid ambiguity.
  2. [Section 5.2, Table 5] Table 5 lists 'Loop Ordering: Exhaustive,' but Section 5.2 actually proposes two strategies (iterative and softmax-based gradient descent), and Section 6.2 uses iterative optimization. The table should be updated to reflect the implemented search method.
  3. [Section 6.5] The 'real hardware' experiments evaluate latency with FireSim but still estimate energy with Timeloop/Accelergy. This is fine, but it should be stated in one place, e.g., in the abstract or at the start of Section 6.5, to avoid overclaiming that the RTL study covers the full EDP.
  4. [Section 6.5.2] The DNN correction model is trained for 50,000 epochs on 1,567 samples. It would be helpful to report early stopping, regularization, or training/validation curves, because at that epoch count overfitting is a plausible concern even with a small model.
  5. [Section 7, Conclusion] The conclusion says DOSA uses 'a similar number of samples,' which repeats the sample-equivalence issue. Please rephrase to reflect the corrected comparison proposed in the major comments.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: DOSA's differentiable model is validated against Timeloop and final designs are evaluated with independent Timeloop/FireSim evaluation, so the central claims do not reduce to their inputs.

full rationale

The paper's derivation chain is self-contained. The differentiable model (Section 4) is a closed-form transcription of Timeloop's analytical relationships; validating it against Timeloop (Section 4.6, Fig. 4) is a consistency check, not a circular reuse of the target result. The optimization objective is this differentiable model, but the reported EDP improvements are evaluated with Timeloop for Gemmini-TL (Section 6.3, Fig. 7-8) and with FireSim RTL simulation for Gemmini-RTL (Section 6.5, Fig. 12), i.e., independent evaluators outside the model's own loss. The DNN correction model (Section 4.7) is trained on FireSim data from training workloads and evaluated on held-out target workloads and DOSA-generated mappings (Section 6.5.2, Fig. 10-11), so its predictions are not fitted to the reported outcomes. Hardware parameters are derived from mappings by the explicit minimal-parameterization max operation (Eq. 1-5); this is a disclosed search-space assumption, not a hidden reuse of the answer. No load-bearing self-citation chain or imported uniqueness theorem is invoked: CoSA is used for initialization and as a comparison baseline, and Timeloop/FireSim are external, widely used tools. The main caveats — the fixed weight-stationary C-K dataflow and bypassing (Section 5.1), the exclusion of pipeline fusion (Section 5.3.4), and the Section 6.3 convention that 'Evaluations done using Timeloop are considered equivalent to evaluations done using DOSA's differentiable model' — affect generalizability and the comparability of sample-efficiency numbers, but they are evaluation/scope limitations, not circular derivations. The paper's quantitative claims therefore do not reduce by construction to their inputs.

Assumptions & free parameters 2 free parameters · 5 assumptions · 0 invented entities

The central claim rests on a small set of domain assumptions: Timeloop as ground truth, the mapping-first hardware parameterization, the roofline latency model, and fixed dataflow. The only fitted model is the DNN correction, whose weights are learned from RTL data. No new physical entities are introduced.

free parameters (2)
  • DNN correction model weights = 5737 parameters
    Trained on 1567 FireSim latency samples (Section 6.5.1) to predict the difference between analytical model and Gemmini-RTL latency; used in Section 4.7 and 6.5.
  • Round-up increment for SRAM sizes = 1 KB
    Design choice in Section 6.1: SRAM sizes rounded up to 1 KB increments; affects resulting hardware configurations.
assumptions (5)
  • domain assumption Timeloop and Accelergy models are accurate proxies for real accelerator behavior
    The central DSE evaluation (Section 6.3-6.4) uses Timeloop as ground truth; Section 6.5 shows analytical-only model has Spearman 0.87 vs Gemmini-RTL, so this assumption is approximate and load-bearing.
  • domain assumption Optimal hardware is the parameter-wise max of per-layer minimal requirements
    Section 4.1: 'we take a parameter-wise max to generate a design that will support all current mappings.' This excludes hardware designs that are not minimal for any single layer's mapping, e.g., to support pipelining or overprovisioning.
  • domain assumption Latency equals the maximum of compute and memory latency, with no overlap between them
    Eq. 12 in Section 4.3. A roofline-style simplification inherited from the intended model; validated only on random mappings.
  • ad hoc to paper Fixed weight-stationary dataflow and fixed bypassing allocation
    Section 5.1 fixes spatial tiling dimensions to C-K weight stationary and allocates one buffer level per tensor precision, reducing the search space; acknowledged in the text.
  • domain assumption The 1567-sample FireSim dataset is representative of the target workloads
    The DNN correction model (Section 6.5.1) is trained on this small dataset; generalization to unseen workloads is evaluated but limited.

how reviews work

0 comments
Cite this review

Pith. "Pith review of DOSA: Differentiable Model-Based One-Loop Search for DNN Accelerators." pith.science (2026). https://pith.science/paper/SQZAYXRO

@misc{pith2026250910702,
  author       = {Pith},
  title        = {Pith review of: DOSA: Differentiable Model-Based One-Loop Search for DNN Accelerators},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/SQZAYXRO}},
  note         = {Machine review of arXiv:2509.10702}
}
read the original abstract

In the hardware design space exploration process, it is critical to optimize both hardware parameters and algorithm-to-hardware mappings. Previous work has largely approached this simultaneous optimization problem by separately exploring the hardware design space and the mapspace - both individually large and highly nonconvex spaces - independently. The resulting combinatorial explosion has created significant difficulties for optimizers. In this paper, we introduce DOSA, which consists of differentiable performance models and a gradient descent-based optimization technique to simultaneously explore both spaces and identify high-performing design points. Experimental results demonstrate that DOSA outperforms random search and Bayesian optimization by 2.80x and 12.59x, respectively, in improving DNN model energy-delay product, given a similar number of samples. We also demonstrate the modularity and flexibility of DOSA by augmenting our analytical model with a learned model, allowing us to optimize buffer sizes and mappings of a real DNN accelerator and attain a 1.82x improvement in energy-delay product.

Figures

Figures reproduced from arXiv: 2509.10702 by the authors.

Figure 1
Figure 1. Hardware-first, two-loop (left) and mapping-first, [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. An architecture diagram of DOSA. accelerator design and mapping while significantly reducing the time and costs associated with DNN accelerator DSE. 3.1 Problem Setup 3.1.1 Target Workloads. DOSA targets accelerator DSE for com￾plete DNN models, which comprise both matrix multiplication and convolution layers. To express these layers, we use seven dimen￾sions: 𝑅 (weight height), 𝑆 (weight width), 𝑃 (output activatio… view at source ↗
Figure 4
Figure 4. Error of DOSA differentiable model prediction with respect to Timeloop for 100 random Gemmini configurations, 73 [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗
Figures from the paper (9 more)
Figure 5
Figure 5. Figure 5: Energy and latency prediction flow different loop [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]
Figure 6
Figure 6. Figure 6: Comparison of no loop ordering optimization by [PITH_FULL_IMAGE:figures/full_fig_p009_6.png]
Figure 7
Figure 7. Figure 7: DOSA EDP optimization of Gemmini-TL on 4 distinct workloads, versus baselines. Each line represents the mean [PITH_FULL_IMAGE:figures/full_fig_p010_7.png]
Figure 8
Figure 8. Figure 8: Energy-delay product (EDP) of baseline accelerators, compared to DOSA-optimized Gemmini-TL. Bar labels represent [PITH_FULL_IMAGE:figures/full_fig_p010_8.png]
Figure 9
Figure 9. Figure 9: DOSA improves performance under a constant map [PITH_FULL_IMAGE:figures/full_fig_p011_9.png]
Figure 10
Figure 10. Figure 10: Accuracy of Gemmini-RTL latency models on test split of random mappings (training workloads from Table 6, [PITH_FULL_IMAGE:figures/full_fig_p012_10.png]
Figure 11
Figure 11. Figure 11: Accuracy of Gemmini-RTL latency modeling on mappings generated using DOSA. These are mappings for the target [PITH_FULL_IMAGE:figures/full_fig_p012_11.png]
Figure 12
Figure 12. Figure 12: DOSA optimization of Gemmini-RTL using various [PITH_FULL_IMAGE:figures/full_fig_p013_12.png]
Figure 7
Figure 7. Figure 7: fig7.sh [PITH_FULL_IMAGE:figures/full_fig_p016_7.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

55 extracted references · 5 linked inside Pith

  1. [1]

    Martín Abadi, Paul Barham, Jianmin Chen, Z. Chen, Andy Davis, Jeffrey Dean, Matthieu Devin, Sanjay Ghemawat, Geoffrey Irving, Michael Isard, Manjunath Kudlur, Josh Levenberg, Rajat Monga, Sherry Moore, Derek Gordon Murray, Benoit Steiner, Paul A. Tucker, Vijay Vasudevan, Pete Warden, Martin Wicke, Yuan Yu, and Xiaoqiang Zhang. 2016. TensorFlow: a system f...

    2016

  2. [2]

    Amazon. 2018. AWS Inferentia: High Performance Machine Learning Inference Chip. https://aws.amazon.com/machine-learning/inferentia/

    2018

  3. [3]

    Tianqi Chen, Mu Li, Yutian Li, Min Lin, Naiyan Wang, Minjie Wang, Tianjun Xiao, Bing Xu, Chiyuan Zhang, and Zheng Zhang. 2015. Mxnet: A flexible and efficient machine learning library for heterogeneous distributed systems. arXiv preprint arXiv:1512.01274 (2015)

    arXiv 2015

  4. [4]

    Yan, Haichen Shen, Meghan Cowan, Leyuan Wang, Yuwei Hu, Luis Ceze, Carlos Guestrin, and Arvind Krishnamurthy

    Tianqi Chen, Thierry Moreau, Ziheng Jiang, Lianmin Zheng, Eddie Q. Yan, Haichen Shen, Meghan Cowan, Leyuan Wang, Yuwei Hu, Luis Ceze, Carlos Guestrin, and Arvind Krishnamurthy. 2018. TVM: An automated end-to-end optimizing compiler for deep learning. In 13th USENIXSymposium on Operating Systems Design and Implementation ({OSDI} 18). 578–594

    2018

  5. [5]

    Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2018. BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding. arXiv preprint arXiv:1810.04805 (2018)

    arXiv 2018

  6. [6]

    Hasan Genc, Seah Kim, Alon Amid, Ameer Haj-Ali, Vighnesh Iyer, Pranav Prakash, Jerry Zhao, Daniel Grubb, Harrison Liew, Howard Mao, Albert Ou, Colin Schmidt, Samuel Steffl, John Wright, Ion Stoica, Jonathan Ragan-Kelley, Krste Asanovic, Borivoje Nikolic, and Yakun Sophia Shao. 2021. Gemmini: En- abling Systematic Deep-Learning Architecture Evaluation via ...

    arXiv 2021

  7. [7]

    Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde- Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio

    Ian J. Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde- Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. 2014. Generative Adversarial Networks. arXiv:1406.2661 [stat.ML]

    arXiv 2014

  8. [8]

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. 2016. Deep Residual Learning for Image Recognition. In Proceedings of the Conference on Computer Vision and Pattern Recognition (CVPR)

    2016

Show all 55 references
  1. [9]

    Kartik Hegde, Po-An Tsai, Sitao Huang, Vikas Chandra, Angshuman Parashar, and Christopher W Fletcher. 2021. Mind mappings: enabling efficient algorithm- accelerator mapping space search. In Proceedings of the International Conference on Architectural Support for Programming La...

    2021

  2. [10]

    Qijing Huang, Charles Hong, John Wawrzynek, Mahesh Subedar, and Yakun Sophia Shao. 2022. Learning A Continuous and Reconstructible Latent Space for Hardware Accelerator Design. In Proceedings of the International Sym- posium on Performance Analysis of Systems and Software (ISPASS)

    2022

  3. [11]

    Qijing Huang, Minwoo Kang, Grace Dinh, Thomas Norell, Aravind Kalaiah, James Demmel, John Wawrzynek, and Yakun Sophia Shao. 2021. CoSA: Schedul- ing by Constrained Optimization for Spatial Accelerators. In Proceedings of the International Symposium on Computer Architecture (ISCA)

    2021

  4. [12]

    Norman P. Jouppi, Cliff Young, Nishant Patil, David Patterson, Gaurav Agrawal, Raminder Bajwa, Sarah Bates, Suresh Bhatia, Nan Boden, Al Borchers, Rick Boyle, Pierre luc Cantin, Clifford Chao, Chris Clark, Jeremy Coriell, Mike Daley, Matt Dau, Jeffrey Dean, Ben Gelb, Tara Vazi...

    2023

  5. [13]

    Sheng-Chun Kao and Tushar Krishna. 2020. GAMMA: Automating the HW Mapping of DNN Models on Accelerators via Genetic Algorithm. In Proceedings of the International Conference on Computer-Aided Design (ICCAD)

    2020

  6. [14]

    Sheng-Chun Kao, Angshuman Parashar, Po-An Tsai, and Tushar Krishna. 2022. Demystifying Map Space Exploration for NPUs. In 2022 IEEE International Sym- posium on Workload Characterization (IISWC). 269–281. https://doi.org/10.1109/ IISWC55918.2022.00031

    2022

  7. [15]

    Sheng-Chun Kao, Michael Pellauer, Angshuman Parashar, and Tushar Kr- ishna. 2022. DiGamma: domain-aware genetic algorithm for HW-mapping co- optimization for DNN accelerators. In 2022 Design, Automation & Test in Europe Conference & Exhibition (DATE). IEEE, 232–237

    2022

  8. [16]

    Sagar Karandikar, Howard Mao, Donggyu Kim, David Biancolin, Alon Amid, Dayeol Lee, Nathan Pemberton, Emmanuel Amaro, Colin Schmidt, Aditya Chopra, Qijing Huang, Kyle Kovacs, Borivoje Nikolic, Randy Katz, Jonathan Bachrach, and Krste Asanovic. 2018. FireSim: FPGA-Accelerated Cy...

    2018

  9. [17]

    John Karro, Greg Kochanski, and Daniel Golovin. 2017. Black box optimization via a bayesian-optimized genetic algorithm. Proc. OPTML (2017), 10th

    2017

  10. [18]

    Sam Kaufman, Phitchaya Phothilimthana, Yanqi Zhou, Charith Mendis, Sudip Roy, Amit Sabne, and Mike Burrows. 2021. A learned performance model for tensor processing units. In Proceedings of Machine Learning and Systems (MLSys)

    2021

  11. [19]

    Fredrik Kjolstad, Shoaib Kamil, Stephen Chou, David Lugato, and Saman Ama- rasinghe. 2017. The tensor algebra compiler. In Proceedings of the International Conference on Object Oriented Programming Systems Languages and Applications (OOPSLA)

    2017

  12. [20]

    Alex Krizhevsky, Ilya Sutskever, and Geoffrey E. Hinton. 2012. Imagenet Classifi- cation with Deep Convolutional Neural Networks. InProceedings of the Conference on Neural Information Processing Systems (NeurIPS)

    2012

  13. [21]

    Aviral Kumar, Amir Yazdanbakhsh, Milad Hashemi, Kevin Swersky, and Sergey Levine. 2021. Data-Driven Offline Optimization for Architecting Hardware Accel- erators. In Workshop on ML for Systems at the Conference on Neural Information Processing Systems (NeurIPS)

    2021

  14. [22]

    Hyoukjun Kwon, Prasanth Chatarasi, Vivek Sarkar, Tushar Krishna, Michael Pellauer, and Angshuman Parashar. 2020. Maestro: A data-centric approach to un- derstand reuse, performance, and hardware cost of dnn mappings. In Proceedings of the International Symposium on Microarchit...

    2020

  15. [23]

    Gary Lauterbach. 2021. The path to successful wafer-scale integration: the Cerebras story. IEEE Micro 41, 6 (2021), 52–57

    2021

  16. [24]

    Rui Li, Yufan Xu, Aravind Sukumaran-Rajam, Atanas Rountev, and P Sadayappan

  17. [25]

    Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dollár. 2017. Focal loss for dense object detection. In Proceedings of the IEEE international conference on computer vision . 2980–2988

    2017

  18. [26]

    Yujun Lin, Mengtian Yang, and Song Han. 2021. NAAS: Neural Accelerator Architecture Search. In Design Automation Conference (DAC)

    2021

  19. [27]

    Liqiang Lu, Naiqing Guan, Yuyue Wang, Liancheng Jia, Zizhang Luo, Jieming Yin, Jason Cong, and Yun Liang. 2021. Tenet: A framework for modeling tensor dataflow based on relation-centric notation. In Proceedings of the International Symposium on Computer Architecture (ISCA)

    2021

  20. [28]

    Linyan Mei, Pouya Houshmand, Vikram Jain, Sebastian Giraldo, and Marian Verhelst. 2021. ZigZag: Enlarging joint architecture-mapping design space explo- ration for DNN accelerators. IEEE Trans. Comput. 70, 8 (2021)

    2021

  21. [29]

    Naveen Muralimanohar, Rajeev Balasubramonian, and Norman Jouppi. 2009. Cacti 6.0: A tool to model large caches. HP Laboratories (01 2009)

    2009

  22. [30]

    Sharan Narang and Greg Diamos. 2017. Baidu DeepBench. GitHub Repository (2017). http://www.github.com/baidu-research/DeepBench

    2017

  23. [31]

    NVIDIA. 2018. TensorRT: https://developer.nvidia.com/tensorrt

    2018

  24. [32]

    NVIDIA. 2020. NVIDIA A100 Tensor Core GPU:. https://www.nvidia.com/en- us/data-center/a100/

    2020

  25. [33]

    OpenAI. 2020. ChatGPT. https://openai.com/blog/chat-gpt-3-launched/. Ac- cessed: April 25, 2023

    2020

  26. [34]

    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. 2019. Timeloop: A Systematic Approach to DNN Accelerator Evaluation. In Proceedings of the Internation...

    2019

  27. [35]

    Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, Alban Desmaison, Andreas Köpf, Edward Yang, Zach DeVito, Martin Raison, Alykhan Tejani, Sasank Chilamkurthy, Benoit Steiner, Lu Fan...

    2019

  28. [36]

    Olaf Ronneberger, Philipp Fischer, and Thomas Brox. 2015. U-net: Convolu- tional networks for biomedical image segmentation. In Medical Image Computing and Computer-Assisted Intervention–MICCAI 2015: 18th International Conference, Munich, Germany, October 5-9, 2015, Proceeding...

    2015

  29. [37]

    Amit Sabne. 2020. Xla: Compiling machine learning for peak performance. (2020)

    2020

  30. [38]

    Chirag Sakhuja, Zhan Shi, and Calvin Lin. 2023. Leveraging Domain Information for the Efficient Automated Design of Deep Learning Accelerators. InInternational Symposium on High-Performance Computer Architectural (HPCA) . IEEE

    2023

  31. [39]

    Yakun Sophia Shao, Brandon Reagen, Gu-Yeon Wei, and David Brooks. 2014. Aladdin: A pre-RTL, power-performance accelerator simulator enabling large design space exploration of customized architectures. In 2014 ACM/IEEE 41st International Symposium on Computer Architecture (ISCA...

    2014

  32. [40]

    Zhan Shi, Chirag Sakhuja, Milad Hashemi, Kevin Swersky, and Calvin Lin. 2020. Using Bayesian Optimization for Hardware/Software Co-Design of Neural Accel- erators. In Workshop on ML for Systems at the Conference on Neural Information Processing Systems (NeurIPS)

    2020

  33. [41]

    Karen Simonyan and Andrew Zisserman. 2014. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556 (2014)

    2014 arXiv

  34. [42]

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Ł ukasz Kaiser, and Illia Polosukhin. 2017. Attention is All you Need. In Proceedings of the Conference on Neural Information Processing Systems (NeurIPS)

    2017

  35. [43]

    Dally, Joel Emer, Stephen W

    Rangharajan Venkatesan, Yakun Sophia Shao, Miaorong Wang, Jason Clemons, Steve Dai, Matthew Fojtik, Ben Keller, Alicia Klinefelter, Nathaniel Pinckney, Priyanka Raina, Yanqing Zhang, Brian Zimmer, William J. Dally, Joel Emer, Stephen W. Keckler, and Brucek Khailany. 2019. Magn...

    2019

  36. [44]

    Irene Wang, Prasenjit Chakraborty, Zi Yu Xue, and Yen Fu Lin. 2022. Evaluation of gem5 for performance modeling of ARM Cortex-R based embedded SoCs. Microprocessors and Microsystems 93 (2022), 104599. https://doi.org/10.1016/j. micpro.2022.104599

    2022

  37. [45]

    Weisstein

    Eric W. Weisstein. [n. d.]. Spearman Rank Correlation Coefficient. ([n. d.]). https://mathworld.wolfram.com/SpearmanRankCorrelationCoefficient.html

  38. [46]

    Samuel Williams, Andrew Waterman, and David Patterson. 2009. Roofline: an insightful visual performance model for multicore architectures. Commun. ACM 52, 4 (2009), 65–76

    2009

  39. [47]

    Emer, and Vivienne Sze

    Yannan Nellie Wu, Joel S. Emer, and Vivienne Sze. 2019. Accelergy: An Architecture-Level Energy Estimation Methodology for Accelerator Designs. In 2019 IEEE/ACM International Conference on Computer-Aided Design (ICCAD) . 1–8. https://doi.org/10.1109/ICCAD45719.2019.8942149

    2019

  40. [48]

    Yannan Nellie Wu, Po-An Tsai, Angshuman Parashar, Vivienne Sze, and Joel S Emer. 2022. Sparseloop: An Analytical Approach To Sparse Tensor Accelerator Modeling. In Proceedings of the International Symposium on Microarchitecture (MICRO). IEEE

    2022

  41. [49]

    Sam Likun Xi, Hans Jacobson, Pradip Bose, Gu-Yeon Wei, and David Brooks. 2015. Quantifying sources of error in McPAT and potential impacts on architectural studies. In 2015 IEEE 21st International Symposium on High Performance Computer Architecture (HPCA). 577–589. https://doi...

    2015

  42. [50]

    Qingcheng Xiao, Size Zheng, Bingzhe Wu, Pengcheng Xu, Xuehai Qian, and Yun Liang. 2021. Hasco: Towards agile hardware and software co-design for tensor computation. In Proceedings of the International Symposium on Computer Architecture (ISCA)

    2021

  43. [51]

    Saining Xie, Ross Girshick, Piotr Dollar, Zhuowen Tu, and Kaiming He. 2017. Aggregated Residual Transformations for Deep Neural Networks. In Proceedings of the Conference on Computer Vision and Pattern Recognition (CVPR)

    2017

  44. [52]

    Xuan Yang, Mingyu Gao, Qiaoyi Liu, Jeff Setter, Jing Pu, Ankita Nayak, Steven Bell, Kaidi Cao, Heonjae Ha, Priyanka Raina, Christos Kozyrakis, and Mark Horowitz. 2020. Interstellar: Using Halide’s Scheduling Language to Analyze DNN Accelerators. In Proceedings of the Internati...

    2020

  45. [53]

    Amir Yazdanbakhsh, Christof Angermueller, Berkin Akin, Yanqi Zhou, Albin Jones, Milad Hashemi, Kevin Swersky, Satrajit Chatterjee, Ravi Narayanaswami, and James Laudon. 2021. Apollo: Transferable Architecture Exploration. arXiv preprint arXiv:2102.01723 (2021)

    2021 arXiv

  46. [54]

    manager” instance), and one f1.2xlarge instance (

    Dan Zhang, Safeen Huda, Ebrahim Songhori, Kartik Prabhu, Quoc Le, Anna Goldie, and Azalia Mirhoseini. 2022. A Full-Stack Search Technique for Domain Optimized Deep Learning Accelerators. In Proceedings of the 27th ACM Interna- tional Conference on Architectural Support for Pro...

    2022

  47. [2021]

    In Proceedings of the International Conference on Architectural Support for Programming Languages and Operation Systems (ASPLOS)

    Analytical characterization and design space exploration for optimization of CNNs. In Proceedings of the International Conference on Architectural Support for Programming Languages and Operation Systems (ASPLOS)

Pith tools

Reviewed August 4, 2026 · model on record in the stance chip above.