Pith. sign in

REVIEW 3 major objections 7 minor 80 references

EdgeMark: An Automation and Benchmarking System for Embedded Artificial Intelligence Tools

T0 review · 3 major / 7 minor · reviewed 2026-08-09 · deepseek-v4-flash

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

Pith's one-line read This paper presents EdgeMark, an open-source system that automates model generation, conversion, deployment, and benchmarking of embedded AI tools, with experiments that guide tool and quantization choices.

desk verdict EdgeMark is a useful, mostly transparent TinyML benchmark, but the TFLM-vs-Ekkono comparison rests on an unvalidated 'minimal changes' assumption that should be fixed before the guidance is taken as settled. read the letter →

arxiv 2502.01700 v1 pith:DUVED37T submitted 2025-02-03 cs.LG

Mohammad Amin Hasanpour , Mikkel Kirkegaard , Xenofon Fafoutis This is my paper · ORCID
classification cs.LG MSC 68T99
keywords TinyMLembeddedAIbenchmarkingautomationsystemquantizationmodeldeploymentmicrocontrollersEdgeMark
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 the many tools for running machine learning on microcontrollers can be compared fairly and efficiently by automating the whole deployment pipeline, and that doing so yields practical guidance for tool and quantization choices. It introduces EdgeMark, an open-source modular system that generates TensorFlow models, converts them through TFLite to tool-specific C/C++ code, flashes them to two microcontroller boards, and reports execution time, flash, RAM, and deployment error. Using thousands of such automated tests, the paper claims to show, among other things, that int8-only quantization is a good default for most cases, that pruning and clustering do not speed up execution on general-purpose cores, and that no single tool wins on every metric. A sympathetic reader would care because these choices currently require time-consuming manual experiments, and the paper's automation plus its open data make the trade-offs measurable and reproducible.

What carries the argument

The load-bearing mechanism is the EdgeMark pipeline itself: modular function blocks (Generate TF Models, Generate Ekkono Models, Convert to TFLite/TFLM/Edge Impulse/eAI Translator, and two board-test modules) chained by a main script and configured by YAML files. For each deployed model it measures execution time as the average of 10 runs, derives flash and RAM from compiler reports by subtracting a base project, computes deployment error against a PC reference, and finds the minimum TFLite arena size with a search algorithm that grows and shrinks a guess by fixed rates until the lower and upper bounds meet. This automation is what makes thousands of comparable measurements across tools, quantization schemes, and compiler optimizations feasible.

What would settle it

Run the same EdgeMark model suite with default settings on a third microcontroller—for example a Cortex-M0 board without an FPU—and compare execution time, flash, and RAM across TFLM, Edge Impulse, and eAI Translator; if int8-only quantization is no longer a good general-purpose default, or if the tool rankings reverse, the paper's headline guidance fails.

Watch

Extended reading notes

Core claim

The central claim, stated on the paper's own terms, is that EdgeMark works: it successfully automates model generation, optimization, conversion, and deployment for TFLM, Edge Impulse, Ekkono, and Renesas eAI Translator, and the benchmarks it produces are trustworthy enough to guide developers. From those benchmarks the paper concludes that dynamic-range quantization should be avoided, that int8-only quantization is a good choice for most models, that pruning and clustering provide no execution-time benefit on general-purpose processors, and that tool choice depends on the metric: Edge Impulse saves flash on the STM board, TFLM is faster on small models, Ekkono is lighter for small regression models, and the vendor-specific eAI Translator is generally best on the Renesas RX65N. The paper also reports surprising RNN results (quantized variants are not consistently smaller or faster) and that an FPU gives a large speedup for float models but little for int8-only ones.

Load-bearing premise

The load-bearing premise is that the two boards (STM NUCLEO-L4R5ZI and Renesas RX65N), default IDE settings, the chosen model families, and the measurement protocol are representative enough that the observed trade-offs generalize to other embedded settings, a premise the paper itself qualifies in Appendix A.4 by stating its board comparison should not be considered comprehensive.

Editorial extensions

If this is right

  • Int8-only quantization is a good default for most cases, especially on larger FC and CNN models, while basic float models are worth keeping only when accuracy is paramount and memory is not constrained.
  • Pruning and clustering provide no meaningful speed or memory benefit on general-purpose microcontroller cores, so developers should reserve sparsity techniques for structured pruning or hardware that explicitly accelerates sparse computation.
  • Edge Impulse's EON-compiled models consume less flash than TFLM on the STM board but TFLM is faster on small models; on the Renesas RX65N, TFLM outperforms Edge Impulse on all three metrics.
  • For small regression models Ekkono is the most efficient choice, while TFLM's int8-only variant becomes better as models grow; on Renesas hardware the vendor-specific eAI Translator is generally the best across execution time, flash, and RAM.
  • RNN deployments on microcontrollers defy simple expectations: int8-only variants were not consistently faster or smaller than basic ones, and GRUs used more flash and RAM than comparable LSTMs despite having fewer parameters.

Reading between the lines

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

  • If EdgeMark's reproducibility claim holds, the same modular pipeline could naturally absorb energy measurement and hardware accelerator support, turning it into a reusable harness for neural architecture search with real on-device metrics rather than estimates.
  • The finding that pruning and clustering do not help on general-purpose MCUs suggests a testable design rule: optimization budgets on such devices should go to quantization, operator fusion, and hardware-specific kernel tuning, not sparsity.
  • The board comparison caveat implies tool rankings may shift on other ARM cores such as Cortex-M0 or Cortex-M7 with DSP/FPU; adding a third board to EdgeMark's deployment modules would directly test how much of the guidance is platform-specific.
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 / 7 minor

Summary. The paper presents a review of embedded AI (eAI)/TinyML toolchains, introduces an open-source automation system called EdgeMark that generates, converts, deploys, and benchmarks ML models on microcontrollers, and reports experimental comparisons of TensorFlow Lite Micro (TFLM), Edge Impulse, Ekkono, and Renesas eAI Translator on two boards (NUCLEO-L4R5ZI and Renesas RX65N). The model zoo spans FC, CNN, RNN, and MLPerf Tiny architectures. The main empirical findings are that int8-only quantization is a good default, unstructured pruning and clustering do not improve runtime or memory on general-purpose MCU cores, and tool choice depends on model size and the metric of interest, with Ekkono efficient for small models, TFLM int8-only better for larger models, and eAI Translator often superior on Renesas hardware. The paper also documents several correctness failures and exclusions, such as MBNet in the Edge Impulse comparison and the basic version of Simple 2 in the RNN experiments.

Significance. If the reported results hold, the paper makes a practical contribution: it provides a modular, open-source benchmarking system with versioned toolchains, a broad range of models, and a reproducible methodology. The authors explicitly ship code, document tool versions, and include experiments across multiple tools and boards, which is a strength. The comparative results for TFLM versus Edge Impulse and TFLM versus eAI Translator are plausible and generally supported by the displayed data. The main reservation concerns the TFLM versus Ekkono comparison, where the paper's own description implies that the Ekkono models are not necessarily architecture-equivalent to the TFLM models; without evidence of equivalence, the crossover conclusion could be an artifact. Several secondary gaps—such as the absence of Renesas quantization plots, the disclosed exclusion of MBNet, and the failure of Simple 2 basic—limit but do not invalidate the broader guidance.

major comments (3)
  1. [5.2.4 with 3.6 and Section 4] The TFLM-versus-Ekkono comparison lacks evidence that the modified Ekkono models are architecture-equivalent to the TFLM models. The paper states in Section 5.2.4 that 'we have slightly changed some models to make them suitable for regression' and that 'the changes are minimal and should not have a noticeable impact on the results.' However, Section 3.6 and the Generate Ekkono Models module restrict Ekkono to FC regression models without dropout or batch normalization, using only Sigmoid, Tanh, and LeakyReLU nonlinearities. The example configuration file in Section 4 uses ReLU, dropout, and batch-norm options, so at least some FC models compared in Fig. 9 must differ from their original classification versions in activation, normalization, or depth. The paper reports no parameter counts, no layer-by-layer diffs, and no ablation showing that these modifications are performance-neutral. Consequently, the crossover in Fig. 9 ('Ekkono more efficient for smaller models, TFLM int8 only for larger') and the associated conclusion in Section 5.2.4 may be an artifact of comparing different architectures. The authors should specify the exact modifications, verify equal parameter counts and comparable layer structure, or run an ablation to demonstrate negligible impact on execution time, flash, and RAM.
  2. [5.2.1 and Figures 3–6] The paper states in Section 5.2.1 that FC and CNN models were 'tested on the two available boards,' but Figures 3–6 present quantization results only for the NUCLEO-L4R5ZI. The Renesas RX65N results for the quantization study are not shown in the manuscript; the text only notes that some models could not be executed successfully on the Renesas board. Since the guidance 'int8 only quantization is a good choice for most cases' is presented as a general conclusion and is later used as a default in other experiments, the lack of Renesas quantization data in the paper itself weakens the support for that generalization. Either include the corresponding Renesas figures or explicitly restrict the claim to the STM board and note the limitation.
  3. [5.2.3 and Appendix A.1] Two exclusions are disclosed but their potential impact on the conclusions is not analyzed. In Section 5.2.3, MBNet from the MLPerf Tiny suite is excluded from the Edge Impulse comparison because it is 'too large in most test cases.' This leaves the MLPerf comparison without its largest image-classification model, which could affect the flash-size and RAM conclusions for that section. In Appendix A.1, the basic version of Simple 2 'failed to execute on the board' and is excluded from the RNN plots; the conclusion that basic RNN variants are faster and use less flash than int8-only variants is based on the models that did execute, and the failure itself is a correctness finding that should be factored into the guidance. The authors should provide sensitivity checks or explicitly hedge the affected conclusions in light of these exclusions.
minor comments (7)
  1. [All figures] Most figures lack error bars or variance information. The text explains that execution-time standard deviation is near zero, but deployment-error and memory metrics are presented without repeat counts or variance; adding error bars or stating the number of repetitions for each metric would improve precision.
  2. [5.2.1] The quantization names are capitalized inconsistently: the list uses 'Dynamic', 'Int8', 'Int8 only', '16x8', '16x8 int only', and 'Float 16', while the figures and text use lowercase 'int8', 'int8 only', and '16x8 int only'. Please standardize the terminology.
  3. [Section 4] The EdgeMark repository is referenced only by a URL; for archival reproducibility, a versioned DOI or a specific commit hash should be cited, as is done for TFLM.
  4. [5.2.4] The sentence 'This is while the int8 only version of TFLM surpases their performance' contains a typo; 'surpases' should be 'surpasses'.
  5. [3.2] The text in Section 3.2 contains the typo 'executation time' in the description of Edge Impulse's hardware estimation feature.
  6. [5.2.1 and Table 2] The paper does not clarify whether the quantization experiments on the Renesas board used the same model set as the STM experiments; a sentence describing which models ran successfully on Renesas would help interpret the missing figures.
  7. [Appendix A.5] The appendix notes that CC-RX memory requirements could not be interpreted, so only execution time is compared. This limitation is disclosed, but it would be useful to state whether this was due to the compiler output format or the eAI Translator integration.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: EdgeMark's claims rest on external tool executions and hardware measurements, not on fitted parameters or author-derived constraints.

full rationale

This paper is an empirical systems and benchmarking study; its central claims (that EdgeMark automates model generation, conversion, deployment, and benchmarking, and that certain tools/quantization choices trade off execution time, flash, and RAM in observed ways) are supported by actual executions of compiled models on two physical boards (NUCLEO-L4R5ZI and Renesas RX65N) using external tools (TFLM, Edge Impulse, Ekkono, eAI Translator) and external IDEs. There is no fitted parameter whose value is later reported as a prediction, no uniqueness theorem imported from the authors' prior work, and no equation that reduces to its own input. The paper's self-citations ([3], the authors' holistic TinyML review, and [57], the authors' pump cavitation study) are used only as contextual references or as an example that traditional ML algorithms are relevant to industrial applications; neither carries a load-bearing step in the benchmark conclusions. The Ekkono comparison's 'minimal' model changes for regression and the Appendix A.4 caveat about non-comprehensive board comparison are methodological/generalizability limitations, not circularity, because the reported measurements remain external observations rather than consequences of the paper's definitions. Consistent with the reviewer guidance, the honest finding is no significant circularity.

Assumptions & free parameters 1 free parameters · 5 assumptions · 0 invented entities

No fitted free parameters are used to make the benchmark claims, and EdgeMark introduces no new theoretical or physical entities. The conclusions rest on the measurement-representativeness and configuration-fairness assumptions listed above. The arena-search constants are hand-chosen but not load-bearing for the main comparisons.

free parameters (1)
  • Arena search growth/shrink rates and resolution = growth 1.25, shrink 0.8, steps 4, resolution 0.5/1/2 kB
    Algorithm 1 constants chosen by hand to find the minimum TFLM arena size. They are implementation details that do not drive the central benchmark conclusions.
assumptions (5)
  • domain assumption Ten repeated executions with the same input give a stable execution-time estimate
    Used in Section 4 metrics; the paper reports standard deviation close to zero but provides no distribution or run-to-run analysis.
  • domain assumption Ten fixed inputs give a representative deployment-error measurement
    Section 4 defines deployment error as the average over 10 inputs; input diversity and sensitivity are not analyzed.
  • domain assumption Subtracting a base project's compiler-reported flash/RAM isolates the model cost
    Section 4 flash and RAM metrics rely on this subtraction being valid across tools and compilers, including the proprietary Renesas compiler where memory reports were unavailable.
  • domain assumption Default IDE settings with maximum compiler optimization form a fair common configuration
    Section 5 states projects were created with default settings except maximum optimization; the fairness of this configuration across toolchains is assumed rather than justified.
  • domain assumption Minimal regression modifications to Ekkono models do not change the comparison
    Section 5.2.4 asserts the changes 'are minimal and should not have a noticeable impact on the results' without supporting measurements.

how reviews work

0 comments
Cite this review

Pith. "Pith review of EdgeMark: An Automation and Benchmarking System for Embedded Artificial Intelligence Tools." pith.science (2026). https://pith.science/paper/DUVED37T

@misc{pith2026250201700,
  author       = {Pith},
  title        = {Pith review of: EdgeMark: An Automation and Benchmarking System for Embedded Artificial Intelligence Tools},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/DUVED37T}},
  note         = {Machine review of arXiv:2502.01700}
}
read the original abstract

The integration of artificial intelligence (AI) into embedded devices, a paradigm known as embedded artificial intelligence (eAI) or tiny machine learning (TinyML), is transforming industries by enabling intelligent data processing at the edge. However, the many tools available in this domain leave researchers and developers wondering which one is best suited to their needs. This paper provides a review of existing eAI tools, highlighting their features, trade-offs, and limitations. Additionally, we introduce EdgeMark, an open-source automation system designed to streamline the workflow for deploying and benchmarking machine learning (ML) models on embedded platforms. EdgeMark simplifies model generation, optimization, conversion, and deployment while promoting modularity, reproducibility, and scalability. Experimental benchmarking results showcase the performance of widely used eAI tools, including TensorFlow Lite Micro (TFLM), Edge Impulse, Ekkono, and Renesas eAI Translator, across a wide range of models, revealing insights into their relative strengths and weaknesses. The findings provide guidance for researchers and developers in selecting the most suitable tools for specific application requirements, while EdgeMark lowers the barriers to adoption of eAI technologies.

Figures

Figures reproduced from arXiv: 2502.01700 by the authors.

Figure 1
Figure 1. EdgeMark modules and their interconnections. [PITH_FULL_IMAGE:figures/full_fig_p014_1.png] view at source ↗
Figure 2
Figure 2. Number of parameters and MACs for FC and CNN models. [PITH_FULL_IMAGE:figures/full_fig_p023_2.png] view at source ↗
Figure 3
Figure 3. Deployment error across various quantization schemes. The models were tested [PITH_FULL_IMAGE:figures/full_fig_p026_3.png] view at source ↗
Figures from the paper (7 more)
Figure 4
Figure 4. Figure 4: Average execution time across various quantization schemes. The models were [PITH_FULL_IMAGE:figures/full_fig_p027_4.png]
Figure 5
Figure 5. Figure 5: Flash memory usage for various quantization schemes. The models were tested [PITH_FULL_IMAGE:figures/full_fig_p028_5.png]
Figure 6
Figure 6. Figure 6: RAM usage across various quantization schemes. The models were tested on [PITH_FULL_IMAGE:figures/full_fig_p029_6.png]
Figure 7
Figure 7. Figure 7: Performance evaluation of basic, pruned, and clustered CNN models on the [PITH_FULL_IMAGE:figures/full_fig_p032_7.png]
Figure 8
Figure 8. Figure 8: Performance evaluation of FC models deployed using Edge Impulse and TFLM [PITH_FULL_IMAGE:figures/full_fig_p033_8.png]
Figure 9
Figure 9. Figure 9: Performance evaluation of Ekkono and TFLM models on the NUCLEO-L4R5ZI [PITH_FULL_IMAGE:figures/full_fig_p035_9.png]
Figure 10
Figure 10. Figure 10: Performance evaluation of CNN models deployed using eAI Translator and [PITH_FULL_IMAGE:figures/full_fig_p037_10.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

80 extracted references · 74 canonical work pages

  1. [1]

    Tsoukas, A

    V. Tsoukas, A. Gkogkidis, E. Boumpa, A. Kakarountas, A review on the emerging technology of tinyml, ACM Computing Surveys 56 (2024) 1–37. doi:10.1145/3661820

    doi:10.1145/3661820 2024

  2. [2]

    Immonen, T

    R. Immonen, T. H¨ am¨ al¨ ainen, Tiny machine learning for resource- constrained microcontrollers, Journal of Sensors 2022 (2022) 1–11. doi: 10.1155/2022/7437023

    work page doi:10.1155/2022/7437023 2022

  3. [3]

    E. Njor, M. A. Hasanpour, J. Madsen, X. Fafoutis, A holistic review of the tinyml stack for predictive maintenance, IEEE Access (2024) 1– 1doi:10.1109/ACCESS.2024.3512860

    work page arXiv 2024

  4. [4]

    Sanchez-Iborra, A

    R. Sanchez-Iborra, A. F. Skarmeta, Tinyml-enabled frugal smart ob- jects: Challenges and opportunities, IEEE Circuits and Systems Maga- zine 20 (2020) 4–18. doi:10.1109/MCAS.2020.3005467

    work page arXiv 2020

  5. [5]

    P. P. Ray, A review on tinyml: State-of-the-art and prospects, Journal of King Saud University - Computer and Information Sciences 34 (2022) 1595–1623. doi:10.1016/j.jksuci.2021.11.019

    doi:10.1016/j.jksuci.2021.11.019 2022

  6. [6]

    S. S. Saha, S. S. Sandha, M. Srivastava, Machine learning for microcontroller-class hardware: A review, IEEE Sensors Journal 22 (2022) 21362–21390. doi:10.1109/JSEN.2022.3210773

    work page arXiv 2022

  7. [7]

    Capogrosso, F

    L. Capogrosso, F. Cunico, D. S. Cheng, F. Fummi, M. Cristani, A ma- chine learning-oriented survey on tiny machine learning, IEEE Access 12 (2024) 23406–23426. doi:10.1109/ACCESS.2024.3365349

    work page arXiv 2024

  8. [8]

    Abadade, A

    Y. Abadade, A. Temouden, H. Bamoumen, N. Benamar, Y. Chtouki, A. S. Hafid, A comprehensive survey on tinyml, IEEE Access 11 (2023) 96892–96922. doi:10.1109/ACCESS.2023.3294111. 48

    work page arXiv 2023

Show all 80 references
  1. [9]

    C. R. Banbury, V. J. Reddi, M. Lam, W. Fu, A. Fazel, J. Holleman, X. Huang, R. Hurtado, D. Kanter, A. Lokhmotov, D. Patterson, D. Pau, J. sun Seo, J. Sieracki, U. Thakker, M. Verhelst, P. Yadav, Benchmark- ing tinyml systems: Challenges and direction, ArXiv (3 2020)

    2020

  2. [10]

    Osman, U

    A. Osman, U. Abid, L. Gemma, M. Perotto, D. Brunelli, Tinyml plat- forms benchmarking, in: S. Saponara, A. De Gloria (Eds.), Applications in Electronics Pervading Industry, Environment and Society, Springer International Publishing, Cham, 2022, pp. 139–148

    2022

  3. [11]

    Wulfert, J

    L. Wulfert, J. K¨ uhnel, L. Krupp, J. Viga, C. Wiede, P. Gembaczka, A. Grabmaier, Aifes: A next-generation edge ai framework, IEEE Trans- actions on Pattern Analysis and Machine Intelligence 46 (2024) 4519–

    2024

  4. [12]

    van Kempen, R

    P. van Kempen, R. Stahl, D. Mueller-Gritschneder, U. Schlichtmann, Mlonmcu: Tinyml benchmarking with fast retargeting, in: Proceedings of the 2023 Workshop on Compilers, Deployment, and Tooling for Edge AI, ACM, 2023, pp. 32–36. doi:10.1145/3615338.3618128

    2023

  5. [13]

    Baciu, J

    V.-E. Baciu, J. Stiens, B. da Silva, Mlino bench: A comprehensive benchmarking tool for evaluating ml models on edge devices, Journal of Systems Architecture 155 (2024) 103262. doi:10.1016/j.sysarc.2 024.103262

    2024 doi

  6. [14]

    David, J

    R. David, J. Duke, A. Jain, V. J. Reddi, N. Jeffries, J. Li, N. Kreeger, I. Nappier, M. Natraj, S. Regev, R. Rhodes, T. Wang, P. Warden, Ten- sorflow lite micro: Embedded machine learning on tinyml systems, Pro- ceedings of Machine Learning and Systems 3 (MLSys 2021) (10 2020)

    2021

  7. [15]

    Hymel, C

    S. Hymel, C. Banbury, D. Situnayake, A. Elium, C. Ward, M. Kel- cey, M. Baaijens, M. Majchrzycki, J. Plunkett, D. Tischler, A. Grande, L. Moreau, D. Maslov, A. Beavis, J. Jongboom, V. J. Reddi, Edge im- pulse: An mlops platform for tiny machine learning (2023). arXiv: 2212.033...

    2023 arXiv

  8. [16]

    URL https://edgeimpulse.com/blog/introducing-eon 49

    Edgeimpulse, Introducing eon: Neural networks in up to 55% less ram and 35% less rom, accessed: 16 January, 2025. URL https://edgeimpulse.com/blog/introducing-eon 49

    2025

  9. [17]

    URL https://tvm.apache.org/docs/topic/microtvm

    Apache TVM, microtvm: Tvm on bare-metal, accessed: 16 January, 2025. URL https://tvm.apache.org/docs/topic/microtvm

    2025

  10. [18]

    Weber, A

    L. Weber, A. Reusch, Tinyml - how tvm is taming tiny, accessed: 16 January, 2025 (6 2020). URL https://tvm.apache.org/2020/06/04/tinyml-how-tvm-is-t aming-tiny

    2025

  11. [19]

    URL https://stm32ai.st.com/stm32-cube-ai

    STMicroelectronics, Stm32cube.ai - stmicroelectronics - stm32 ai, ac- cessed: 16 January, 2025. URL https://stm32ai.st.com/stm32-cube-ai

    2025

  12. [20]

    STMicroelectronics, Microcontrollers stm32cubeai solution, accessed: 16 January, 2025. URL https://www.st.com/content/ccc/resource/sales_and_mark eting/presentation/product_presentation/group0/69/82/bf/ae /5a/8b/40/91/STM32CubeAI_press_pres/files/STM32CubeAI_pres s_pres.pdf/jc...

    2025

  13. [21]

    URL https://www.renesas.com/en/e-ai-development-environme nt-microcontrollers

    Renesas Electronics Corporation, e-ai development environment for mi- crocontrollers, accessed: 16 January, 2025. URL https://www.renesas.com/en/e-ai-development-environme nt-microcontrollers

    2025

  14. [22]

    URL https://www.ekkono.ai

    Ekkono, Edge machine learning and virtual sensors - ekkono solutiona, accessed: 16 January, 2025. URL https://www.ekkono.ai

    2025

  15. [23]

    URL https://www.arm.com/products/silicon-ip-cpu/ethos/arm -nn

    Arm, Arm nn sdk — efficient ml for arm cpus, gpus, & npus - arm, accessed: 16 January, 2025. URL https://www.arm.com/products/silicon-ip-cpu/ethos/arm -nn

    2025

  16. [24]

    URL https://microsoft.github.io/ELL 50

    Microsoft Corporation, The embedded learning library - embedded learning library (ell), accessed: 16 January, 2025. URL https://microsoft.github.io/ELL 50

    2025

  17. [25]

    URL https://www.st.com/en/development-tools/nanoedgeaistu dio.html

    STMicroelectronics, Nanoedgeaistudio - automated machine learning (ml) tool for stm32 developers - stmicroelectronics, accessed: 16 Jan- uary, 2025. URL https://www.st.com/en/development-tools/nanoedgeaistu dio.html

    2025

  18. [26]

    URL https://wiki.st.com/stm32mcu/wiki/AI:NanoEdge_AI_Stud io

    STMicroelectronics, Ai:nanoedge ai studio - stm32mcu, accessed: 16 January, 2025 (6 2024). URL https://wiki.st.com/stm32mcu/wiki/AI:NanoEdge_AI_Stud io

    2025

  19. [27]

    URL https://newsroom.st.com/media-center/press-item.html/n 4592.html

    STMicroelectronics, Stmicroelectronics breaks down barriers to edge ai adoption with free nanoedge ai deployment - st news, accessed: 16 Jan- uary, 2025. URL https://newsroom.st.com/media-center/press-item.html/n 4592.html

    2025

  20. [28]

    URL https://utensor.github.io/website

    uTensor, microtensor, accessed: 16 January, 2025. URL https://utensor.github.io/website

    2025

  21. [29]

    Shelby, utensor and tensor flow announcement — mbed, accessed: 16 January, 2025 (5 2019)

    Z. Shelby, utensor and tensor flow announcement — mbed, accessed: 16 January, 2025 (5 2019). URL https://os.mbed.com/blog/entry/uTensor-and-Tensor-Flo w-Announcement

    2025

  22. [30]

    Lord, Important update on mbed — mbed, accessed: 16 January, 2025

    W. Lord, Important update on mbed — mbed, accessed: 16 January, 2025. URL https://os.mbed.com/blog/entry/Important-Update-on-Mbe d

    2025

  23. [31]

    Paissan, micromind: A toolkit for tinyml research and deployment, accessed: 16 January, 2025

    F. Paissan, micromind: A toolkit for tinyml research and deployment, accessed: 16 January, 2025. URL https://github.com/micromind-toolkit/micromind

    2025

  24. [32]

    URL https://pytorch.org/executorch-overview

    The PyTorch Foundation, Pytorch executorch — pytorch, accessed: 16 January, 2025. URL https://pytorch.org/executorch-overview

    2025

  25. [33]

    URL https://www.imagimob.com 51

    Imagimob AB, Imagimob - edge ai — tinyml — deep learning, accessed: 16 January, 2025. URL https://www.imagimob.com 51

    2025

  26. [34]

    - arm, accessed: 16 January, 2025

    Arm, Omniml inc. - arm, accessed: 16 January, 2025. URL https://www.arm.com/partners/catalog/omnimlinc

    2025

  27. [35]

    Jaiswal, R

    S. Jaiswal, R. K. K. Goli, A. Kumar, V. Seshadri, R. Sharma, Mi- nun: Accurate ml inference on microcontrollers, in: Proceedings of the 24th ACM SIGPLAN/SIGBED International Conference on Languages, Compilers, and Tools for Embedded Systems, ACM, 2023, pp. 26–39. doi:10.1145/3...

    2023

  28. [36]

    Lin, W.-M

    J. Lin, W.-M. Chen, Y. Lin, C. Gan, S. Han, Mcunet: Tiny deep learning on iot devices, Advances in Neural Information Processing Systems 33 (2020)

    2020

  29. [37]

    X. Wang, M. Magno, L. Cavigelli, L. Benini, Fann-on-mcu: An open- source toolkit for energy-efficient neural network inference at the edge of the internet of things, IEEE Internet of Things Journal (11 2019)

    2019

  30. [38]

    URL https://www.ims.fraunhofer.de/en/Business-Unit/Indust ry/Industrial-AI/Artificial-Intelligence-for-Embedded-Sys tems-AIfES.html

    Fraunhofer Institute for Microelectronic Circuits and Systems, Artificial intelligence for embedded systems - fraunhofer ims, accessed: 16 Jan- uary, 2025. URL https://www.ims.fraunhofer.de/en/Business-Unit/Indust ry/Industrial-AI/Artificial-Intelligence-for-Embedded-Sys tems-...

    2025

  31. [39]

    Salerno, tinymlgen: Generate c code for microcontrollers from ten- sorflow models, accessed: 16 January, 2025 (2020)

    S. Salerno, tinymlgen: Generate c code for microcontrollers from ten- sorflow models, accessed: 16 January, 2025 (2020). URL https://github.com/eloquentarduino/tinymlgen

    2020

  32. [40]

    Rusci, A

    M. Rusci, A. Capotondi, L. Benini, Memory-driven mixed low precision quantization for enabling deep network inference on microcontrollers, Proceedings of Machine Learning and Systems (5 2019)

    2019

  33. [41]

    C. B. Moretti, Neurona: Artificial neural networks for arduino, accessed: 16 January, 2025 (2016). URL https://github.com/moretticb/Neurona

    2016

  34. [42]

    URL https://fastmachinelearning.org/hls4ml/ 52

    Fast Machine Learning Lab, hls4ml 1.0.0 documentation, accessed: 16 January, 2025. URL https://fastmachinelearning.org/hls4ml/ 52

    2025

  35. [43]

    URL https://siliconlabs.github.io/mltk

    Silicon Labs, Silicon labs machine learning toolkit (mltk) — mltk 0.20.0 documentation, accessed: 16 January, 2025. URL https://siliconlabs.github.io/mltk

    2025

  36. [44]

    URL https://github.com/microsoft/nni

    Microsoft, Neural network intelligence, accessed: 16 January, 2025 (12 2021). URL https://github.com/microsoft/nni

    2025

  37. [45]

    URL https://developer.apple.com/documentation/coreml

    Apple Inc., Core ml — apple developer documentation, accessed: 16 January, 2025. URL https://developer.apple.com/documentation/coreml

    2025

  38. [46]

    C. Lv, C. Niu, R. Gu, X. Jiang, Z. Wang, B. Liu, Z. Wu, Q. Yao, C. Huang, P. Huang, T. Huang, H. Shu, J. Song, B. Zou, P. Lan, G. Xu, F. Wu, S. Tang, F. Wu, G. Chen, Walle: An end-to-end, general- purpose, and large-scale production system for device-cloud collabo- rative mach...

    2022

  39. [47]

    URL https://ai.meta.com/tools/glow

    Meta, Glow, accessed: 16 January, 2025. URL https://ai.meta.com/tools/glow

    2025

  40. [48]

    URL https://www.nxp.com/design/design-center/software/eiq -ml-development-environment:EIQ

    NXP Semiconductors, eiq ml software development environment — nxp semiconductors, accessed: 16 January, 2025. URL https://www.nxp.com/design/design-center/software/eiq -ml-development-environment:EIQ

    2025

  41. [49]

    URL https://onnxruntime.ai

    Microsoft, Onnx runtime, accessed: 16 January, 2025. URL https://onnxruntime.ai

    2025

  42. [50]

    C. N. Coelho, A. Kuusela, H. Zhuang, T. Aarrestad, V. Loncar, J. Ngadi- uba, M. Pierini, S. Summers, Ultra low-latency, low-area inference accel- erators using heterogeneous deep quantization with qkeras and hls4ml, arXiv preprint arXiv:2006.10159 (2020) 108

    2020 arXiv

  43. [51]

    URL https://larq.dev 53

    Plumerai, Larq, accessed: 16 January, 2025. URL https://larq.dev 53

    2025

  44. [52]

    URL https://onnc.ai

    Skymizer Taiwan Inc., Onnc, accessed: 16 January, 2025. URL https://onnc.ai

    2025

  45. [53]

    URL https://latentai.com

    Latent AI, Latent ai - find your best model faster, accessed: 16 January, 2025. URL https://latentai.com

    2025

  46. [54]

    URL https://plumerai.com/

    Plumerai, Plumerai, accessed: 16 January, 2025. URL https://plumerai.com/

    2025

  47. [55]

    URL https://greenwaves-technologies.com/manuals_gap9/gap9_ sdk_doc/html/source/tools/nntool

    GreenWaves Technologies, Nntool — gap sdk documentation, accessed: 16 January, 2025. URL https://greenwaves-technologies.com/manuals_gap9/gap9_ sdk_doc/html/source/tools/nntool

    2025

  48. [56]

    Burrello, A

    A. Burrello, A. Garofalo, N. Bruschi, G. Tagliavini, D. Rossi, F. Conti, Dory: Automatic end-to-end deployment of real-world dnns on low-cost iot mcus, IEEE Transactions on Computers (2021) 1 doi:10.1109/TC .2021.3066883

    2021

  49. [57]

    M. A. Hasanpour, R. Engholm, X. Fafoutis, Pump cavitation detection with machine learning: A comparative study of svm and deep learn- ing, in: 2024 IEEE Annual Congress on Artificial Intelligence of Things (AIoT), 2024, pp. 219–225. doi:10.1109/AIoT63253.2024.00050

    2024

  50. [58]

    Morawiec, sklearn-porter, transpile trained scikit-learn estimators to C, Java, JavaScript and others

    D. Morawiec, sklearn-porter, transpile trained scikit-learn estimators to C, Java, JavaScript and others. URL https://github.com/nok/sklearn-porter

  51. [59]

    Morawiec, weka-porter: Transpile trained decision trees from weka to c, java or javascript., accessed: 16 January, 2025 (2017)

    D. Morawiec, weka-porter: Transpile trained decision trees from weka to c, java or javascript., accessed: 16 January, 2025 (2017). URL https://github.com/nok/weka-porter

    2017

  52. [60]

    Titov, I

    N. Titov, I. Zeigerman, V. Yershov, m2cgen: Transform ml models into a native code (java, c, python, go, javascript, visual basic, c#, r, pow- ershell, php, dart, haskell, ruby, f#, rust) with zero dependencies, ac- cessed: 16 January, 2025. URL https://github.com/BayesWitness...

    2025

  53. [61]

    Salerno, micromlgen: Generate c code for microcontrollers from python’s sklearn classifiers, accessed: 16 January, 2025

    S. Salerno, micromlgen: Generate c code for microcontrollers from python’s sklearn classifiers, accessed: 16 January, 2025. URL https://github.com/eloquentarduino/micromlgen

    2025

  54. [62]

    L. T. da Silva, V. M. A. Souza, G. E. A. P. A. Batista, An open-source tool for classification models in resource-constrained hardware, IEEE Sensors Journal 22 (2022) 544–554. doi:10.1109/JSEN.2021.3128130

    2022

  55. [63]

    URL https://docs.qeexo.com/guides/userguides

    Qeexo AutoML, Automl by tdk sensei help center, accessed: 16 January, 2025. URL https://docs.qeexo.com/guides/userguides

    2025

  56. [64]

    Nordby, emlearn: Machine learning inference engine for microcon- trollers and embedded devices (3 2019)

    J. Nordby, emlearn: Machine learning inference engine for microcon- trollers and embedded devices (3 2019). doi:10.5281/zenodo.2589394. URL https://doi.org/10.5281/zenodo.2589394

    2019 doi

  57. [65]

    Howard, M

    A. Howard, M. Sandler, G. Chu, L.-C. Chen, B. Chen, M. Tan, W. Wang, Y. Zhu, R. Pang, V. Vasudevan, et al., Searching for mo- bilenetv3, in: Proceedings of the IEEE/CVF international conference on computer vision, 2019, pp. 1314–1324

    2019

  58. [66]

    H. Cai, C. Gan, T. Wang, Z. Zhang, S. Han, Once-for-all: Train one network and specialize it for efficient deployment, arXiv preprint arXiv:1908.09791 (2019)

    2019 arXiv

  59. [67]

    Gambella, A

    M. Gambella, A. Falcetta, M. Roveri, Cnas: Constrained neural ar- chitecture search, in: 2022 IEEE International Conference on Systems, Man, and Cybernetics (SMC), IEEE, 2022, pp. 2918–2923

    2022

  60. [68]

    Lomurno, S

    E. Lomurno, S. Mariani, M. Monti, M. Matteucci, Pomonag: Pareto- optimal many-objective neural architecture generator, arXiv preprint arXiv:2409.20447 (2024)

    2024 arXiv

  61. [69]

    X. Liu, W. Liu, L.-N. Wang, G. Zhong, Deep architecture compression with automatic clustering of similar neurons, in: Pattern Recognition and Computer Vision: 4th Chinese Conference, PRCV 2021, Beijing, China, October 29–November 1, 2021, Proceedings, Part IV 4, Springer, 2021...

    2021

  62. [70]

    Hinton, O

    G. Hinton, O. Vinyals, J. Dean, et al., Distilling the knowledge in a neural network, arXiv preprint arXiv:1503.02531 2 (7) (2015). 55

    2015 arXiv

  63. [71]

    N. P. Ghanathe, S. Wilton, T-recx: Tiny-resource efficient convolutional neural networks with early-exit, in: Proceedings of the 20th ACM In- ternational Conference on Computing Frontiers, 2023, pp. 123–133

    2023

  64. [72]

    J. Lin, L. Zhu, W.-M. Chen, W.-C. Wang, C. Gan, S. Han, On-device training under 256kb memory, in: S. Koyejo, S. Mohamed, A. Agarwal, D. Belgrave, K. Cho, A. Oh (Eds.), Advances in Neural Information Processing Systems, Vol. 35, Curran Associates, Inc., 2022, pp. 22941– 22954....

    2022

  65. [73]

    Pavan, E

    M. Pavan, E. Ostrovan, A. Caltabiano, M. Roveri, Tybox: An automatic design and code generation toolbox for tinyml incremental on-device learning, ACM Transactions on Embedded Computing Systems 23 (3) (2024) 1–27

    2024

  66. [74]

    H. Ren, D. Anicic, T. A. Runkler, Tinyol: Tinyml with online-learning on microcontrollers, in: 2021 international joint conference on neural networks (IJCNN), IEEE, 2021, pp. 1–8

    2021

  67. [75]

    Karpathy, char-rnn, https://github.com/karpathy/char-rnn , accessed: 16 January, 2025 (2015)

    A. Karpathy, char-rnn, https://github.com/karpathy/char-rnn , accessed: 16 January, 2025 (2015)

    2015

  68. [76]

    URL https://ai.google.dev/edge/litert/models/post_training _quant

    Google - The AI Edge Authors, Post-training dynamic range quantiza- tion, accessed: 16 January, 2025. URL https://ai.google.dev/edge/litert/models/post_training _quant

    2025

  69. [77]

    URL https://ai.google.dev/edge/litert/models/post_training _quant

    Google - The AI Edge Authors, Post-training integer quantization, ac- cessed: 16 January, 2025. URL https://ai.google.dev/edge/litert/models/post_training _quant

    2025

  70. [78]

    URL https://ai.google.dev/edge/litert/models/post_training _integer_quant_16x8

    Google - The AI Edge Authors, Post-training integer quantization with int16 activations, accessed: 16 January, 2025. URL https://ai.google.dev/edge/litert/models/post_training _integer_quant_16x8

    2025

  71. [79]

    56 URL https://ai.google.dev/edge/litert/models/post_training _float16_quant 57

    Google - The AI Edge Authors, Post-training float16 quantization, ac- cessed: 16 January, 2025. 56 URL https://ai.google.dev/edge/litert/models/post_training _float16_quant 57

    2025

  72. [4533]

    doi:10.1109/TPAMI.2024.3355495

    2024

Pith tools

Reviewed August 9, 2026 · model on record in the stance chip above.