CMix-NN: Mixed Low-Precision CNN Library for Memory-Constrained Edge Devices
Open accessAbstract
Low-precision integer arithmetic is a necessary ingredient for enabling Deep Learning inference on tiny and resource-constrained IoT edge devices. This brief presents CMix-NN, a flexible open-sourceCMix-NN is available at https://github.com/EEESlab/CMix-NN. mixed low-precision (independent tensors quantization of weight and activations at 8, 4, 2 bits) inference library for low bitwidth Quantized Networks. CMix-NN efficiently supports both Per-Layer and Per-Channel quantization strategies of weights and activations. Thanks to CMix-NN, we deploy on an STM32H7 microcontroller a set of Mobilenet family networks with the largest input resolutions (224×224) and higher accuracies (up to 68% Top1) when compressed with a mixed low precision technique, achieving up to +8% accuracy improvement concerning any other published solution for MCU devices.
The claims and statistics of this paper haven’t been checked yet.
Run the check — free, no account needed. It reads the open-access text and checks the sources it cites.
Integrity signals
Registry
Cites
Statistics
Claims the paper makes
The claims and statistics of this paper haven’t been checked yet.
Run the check — free, no account needed. It reads the open-access text and checks the sources it cites.
Claims other papers make about this one
Each is checked against this paper’s own text: does it say what the citing paper says it does?
The CMix-NN library is deployed in the STM32H7 microcontroller with a 224x224 resolution capability and higher accuracy (65% Top1).
“The library is deployed in the STM32H7 microcontroller with a 224x224 resolution capability and higher accuracy (65% Top1).”
From A review on TinyML: State-of-the-art and prospects (2022) · cites this paper as Capotondi et al., 2020
The paper reports deploying CMix-NN MobileNet models on an STM32H7 with 224×224 inputs (abstract; IV. Experimental Results). But it reports accuracy of up to 68% Top-1, not the claimed 65% (abstract; B. CMix-NN Network-Level Comparison). (LLM, conf=0.88)
Experimental — accuracy not yet measuredCapotondi et al. (2020) presented CMix-NN to allow flexible and independent tensor quantization for 2, 4, and 8 bits weights.
“This work presents the CMix-NN to allow flexible and independent tensor quantization for 2, 4, 8 bits weights.”
From A review on TinyML: State-of-the-art and prospects (2022) · cites this paper as Capotondi et al., 2020
Self-check
Percentages, sums and reported differences re-computed from the paper’s own numbers.
The self-check needs the paper’s full text.
Cites
Low-bit Quantization of Neural Networks for Efficient Inference
Choukroun, Y., Kravchik, E., Yang, F., & Kisilev, P. (2019) · 2019 IEEE/CVF International Conference on Computer Vision Workshop (ICCVW) · doi:10.1109/iccvw.2019.00363 · Cited by 404 · see citing works
Flagged first, then by citations
Research works printed but not linked: 26
- C.-J. Wu et al., "Machine learning at Facebook: Understanding inference at the edge," in Proc. IEEE Int. Symp. High Perform. Comput. Architect. (HPCA), 2019, pp. 331-344.
- A. Canziani, E. Culurciello, and A. Paszke, "Evaluation of neural network architectures for embedded systems," in Proc. IEEE Int. Symp. Circuits Syst. (ISCAS), 2017, pp. 1-4.
- J. Choi, Z. Wang, S. Venkataramani, P. I.-J. Chuang, V. Srinivasan, and K. Gopalakrishnan, "Pact: Parameterized clipping activation for quan- tized neural networks," 2018. [Online]. Available: arXiv:1805.06085. · arXiv
- M. Nagel, M. van Baalen, T. Blankevoort, and M. Welling, "Data-free quantization through weight equalization and bias correction," 2019. [Online]. Available: arXiv:1906.04721. · arXiv
- S. R. Jain, A. Gural, M. Wu, and C. Dick, "Trained uniform quantiza- tion for accurate and efficient neural network inference on fixed-point hardware," 2019. [Online]. Available: arXiv:1903.08066. · arXiv
- S. K. Esser, J. L. McKinstry, D. Bablani, R. Appuswamy, and D. S. Modha, "Learned step size quantization," 2019. [Online]. Available: arXiv:1902.08153. · arXiv
- R. Krishnamoorthi, "Quantizing deep convolutional networks for efficient inference: A whitepaper," 2018. [Online]. Available: arXiv:1806.08342. · arXiv
- Z.-G. Liu and M. Mattina, "Learning low-precision neural networks without straight-through estimator (STE)," 2019. [Online]. Available: arXiv:1903.01061. · arXiv
- Y. Umuroglu and M. Jahre, "Streamlined deployment for quantized neural networks," 2017. [Online]. Available: arXiv:1709.04060. · arXiv
- H. Gao, W. Tao, D. Wen, T.-W. Chen, K. Osa, and M. Kato, "IFQ- Net: Integrated fixed-point quantization networks for embedded vision," in Proc. IEEE Conf. Comput. Vis. Pattern Recognit. Workshops, 2018, pp. 607-615.
- M. Rusci, A. Capotondi, F. Conti, and L. Benini, "Work-in-progress: Quantized NNS as the definitive solution for inference on low-power ARM MCUs?" in Proc. Int. Conf. Hardw./Softw. Codesign Syst. Synth. (CODES+ ISSS), 2018, pp. 1-2.
- B. Jacob et al., "Quantization and training of neural networks for effi- cient integer-arithmetic-only inference," in Proc. IEEE Conf. Comput. Vis. Pattern Recognit., 2018, pp. 2704-2713.
- M. Rusci, A. Capotondi, and L. Benini, "Memory-driven mixed low precision quantization for enabling deep network inference on micro- controllers," 2019. [Online]. Available: arXiv:1905.13082. · arXiv
- Y. LeCun, K. Kavukcuoglu, and C. Farabet, "Convolutional networks and applications in vision," in Proc. IEEE Int. Symp. Circuits Syst., May 2010, pp. 253-256.
- M. Blott et al., "FINN-R: An end-to-end deep-learning framework for fast exploration of quantized neural networks," ACM Trans. Reconfig. Technol. Syst. (TRETS), vol. 11, no. 3, p. 16, 2018.
- V. Gokhale, A. Zaidy, A. X. M. Chang, and E. Culurciello, "Snowflake: An efficient hardware accelerator for convolutional neural networks," in Proc. IEEE Int. Symp. Circuits Syst. (ISCAS), 2017, pp. 1-4.
- L. Bai, Y. Zhao, and X. Huang, "A CNN accelerator on FPGA using depthwise separable convolution," IEEE Trans. Circuits Syst. II, Exp. Briefs, vol. 65, no. 10, pp. 1415-1419, Oct. 2018.
- J. Lee, J. K. Eshraghian, K. Cho, and K. Eshraghian, "Adaptive precision cnn accelerator using radix-x parallel connected memristor crossbars," 2019. [Online]. Available: arXiv:1906.09395. · arXiv
- E. Wang et al., "Intel math kernel library," in High-Performance Computing on the Intel Xeon Phi TM . Cham, Switzerland: Springer, 2014, pp. 167-188.
- M. Cowan, T. Moreau, T. Chen, and L. Ceze, "Automating generation of low precision deep learning operators," 2018. [Online]. Available: arXiv:1810.11066. · arXiv
- A. Tulloch and Y. Jia, "High performance ultra-low-precision convolu- tions on mobile devices," 2017. [Online]. Available: arXiv:1712.02427. · arXiv
- L. Lai, N. Suda, and V. Chandra, "CMSIS-NN: Efficient neu- ral network kernels for arm cortex-M CPUs," 2018. [Online]. Available: arXiv:1801.06601. · arXiv
- Y. Zhang, N. Suda, L. Lai, and V. Chandra, "Hello edge: Keyword spot- ting on microcontrollers," 2017. [Online]. Available: arXiv:1711.07128. · arXiv
- A. Chowdhery, P. Warden, J. Shlens, A. Howard, and R. Rhodes, "Visual wake words dataset," 2019. [Online]. Available: arXiv:1906.05721. · arXiv
- A. Garofalo, M. Rusci, F. Conti, D. Rossi, and L. Benini, "Pulp- NN: Accelerating quantized neural networks on parallel ultra-low-power RISC-V processors," 2019. [Online]. Available: arxiv:1908.11263. · arXiv
- K. Wang, Z. Liu, Y. Lin, J. Lin, and S. Han, "HAQ: Hardware-aware automated quantization," 2018. [Online]. Available: arXiv:1811.08886. · arXiv
Cited by
Sorted by citations
A review on TinyML: State-of-the-art and prospects
Ray, P. P. (2022) · Journal of King Saud University - Computer and Information Sciences · doi:10.1016/j.jksuci.2021.11.019 · Cited by 460 · see citing works
“Low-precision integer arithmetic is recently tested in a similar mixed deep neural network aware scenario for efficient deployment in the IoT enabled edge devices (Capotondi et al., 2020).”
TinyML-Enabled Frugal Smart Objects: Challenges and Opportunities
Sanchez-Iborra, R., & Skarmeta, A. F. (2020) · IEEE Circuits and Systems Magazine · doi:10.1109/mcas.2020.3005467 · Cited by 295 · see citing works
Alajlan, N. N., & Ibrahim, D. M. (2022) · Micromachines · doi:10.3390/mi13060851 · Cited by 224 · see citing works
“[36] ImageNET Dataset Set of 1000 object categories contains internal and leaf nodes.”
- Authors
- Capotondi, A., Rusci, M., Fariselli, M., & Benini, L.
- Affiliations
- University of Modena and Reggio Emilia (IT)
- University of Bologna (IT)
- Venue
- IEEE Transactions on Circuits and Systems II: Express Briefs
- Published
- 2020-05
- Volume
- 67
- Issue
- 5
- Pages
- 871-875
- ISSN
- 1549-7747 (ISSN-L)
- 1558-3791
- OpenAlex
- W3013186616
- Cited by
- 135
- References
- 1
- Citations by year
- 2026: 9 · 2025: 15 · 2024: 23 · 2023: 31 · 2022: 25 · 2021: 23 · 2020: 9
- FWCI
- 8.56
- Type
- journal article
- Language
- en
- Open access
- green
- hdl.handle.net
- Licence
- Locations
- IEEE Transactions on Circuits & Systems II Express Briefs (publishedVersion)
- IRIS UNIMORE (University of Modena and Reggio Emilia) (submittedVersion, cc-by)
- Archivio istituzionale della ricerca (Alma Mater Studiorum Università di Bologna) (submittedVersion, other-oa)
- Funders
- EU ECSEL-H2020 AI4DI Project (826060)
- Topics
- Physical Sciences › Computer Science › Computer Vision and Pattern Recognition › Advanced Neural Network Applications
- Physical Sciences › Computer Science › Computer Networks and Communications › IoT and Edge/Fog Computing
- Physical Sciences › Engineering › Electrical and Electronic Engineering › Low-power high-performance VLSI design
- Keywords
- integer arithmetic, convolutional neural networks, edge inference, IoT devices, MobileNet, memory-constrained devices
About these checks
Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.
Statistics and arithmetic are recomputed, not judged.
Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.
What counts as a claim ›How we measure confidence ›How we check this ›
Are you an author, or do you see a problem with this page? Request a review
