{
  "title": "Benchmarking Ultra-Low-Power μNPUs",
  "authors": [
    {
      "given": "Josh",
      "family": "Millar",
      "affiliation_ids": [
        1
      ],
      "email": null,
      "orcid": null,
      "corresponding": false
    },
    {
      "given": "Yushan",
      "family": "Huang",
      "affiliation_ids": [
        2
      ],
      "email": null,
      "orcid": null,
      "corresponding": false
    },
    {
      "given": "Sarab",
      "family": "Sethi",
      "affiliation_ids": [
        1
      ],
      "email": null,
      "orcid": null,
      "corresponding": false
    },
    {
      "given": "Hamed",
      "family": "Haddadi",
      "affiliation_ids": [
        1
      ],
      "email": null,
      "orcid": null,
      "corresponding": false
    },
    {
      "given": "Anil",
      "family": "Madhavapeddy",
      "affiliation_ids": [
        3
      ],
      "email": null,
      "orcid": null,
      "corresponding": false
    }
  ],
  "affiliations": [
    {
      "id": 1,
      "text": "Imperial College London, London"
    },
    {
      "id": 2,
      "text": "Imperial College London / Huawei"
    },
    {
      "id": 3,
      "text": "University of Cambridge, Cambridge"
    }
  ],
  "journal": "GetMobile",
  "volume": "30",
  "issue": "2",
  "pages": "11-12",
  "doi": null,
  "published_date": "2026-06",
  "received_date": null,
  "accepted_date": null,
  "abstract": "Microcontrollers (MCUs) are widely used in resource-constrained environments due to their form factor and low cost and power consumption. Performing model inference on MCU-scale hardware improves privacy by running locally, lowers operating costs for model vendors, and eliminates dependence on network connectivity [1, 2]. However, deployments remain constrained by limited memory, throughput, and compute. The growing computational demands of modern models have catalyzed specialized accelerators across the computing spectrum, from data centers to embedded systems. At this resource-constrained end, MCU-scale neural processing units, or µNPUs, have emerged to provide real-time or near-real-time inference within milliwatt-scale power budgets.",
  "keywords": [],
  "funding": [],
  "_opening_blocks": [
    [
      2,
      1
    ],
    [
      2,
      2
    ]
  ],
  "_title_transcribed": "Benchmarking Ultra-Low-Power μNPs",
  "_pdf": {
    "title": "Benchmarking Ultra-Low-Power μNPUs",
    "xmp_title": "ACM SIGMOBILE GetMobile V30 N2 June 2026 NL",
    "title_witness": "the PDF's Info title",
    "doi": null,
    "doi_source": null,
    "outline": [],
    "orcids": [],
    "producer": "Adobe InDesign 21.4 (Macintosh) / Adobe PDF Library 18.0"
  },
  "references": [
    {
      "raw_index": 0,
      "authors": [
        {
          "given": "S.S.",
          "family": "Saha"
        },
        {
          "given": "S.S.",
          "family": "Sandha"
        },
        {
          "given": "M.",
          "family": "Srivastava"
        }
      ],
      "year": "2022",
      "title": "Machine learning for microcontrollerclass hardware: A review",
      "container": "IEEE Sensors Journal",
      "volume": "22",
      "issue": "22",
      "pages": "21362-21390",
      "doi": null,
      "url": null,
      "type": "journal-article"
    },
    {
      "raw_index": 1,
      "authors": [
        {
          "given": "J.",
          "family": "Lin"
        },
        {
          "given": "W.-M.",
          "family": "Chen"
        },
        {
          "given": "Y.",
          "family": "Lin"
        },
        {
          "given": "C.",
          "family": "Gan"
        },
        {
          "given": "S.",
          "family": "Han"
        }
      ],
      "year": "2020",
      "title": "MCUNet: Tiny deep learning on IoT devices",
      "container": "NeurIPS",
      "volume": "33",
      "issue": null,
      "pages": "11711-11722",
      "doi": null,
      "url": null,
      "type": "other"
    },
    {
      "raw_index": 2,
      "authors": [
        {
          "given": "V.",
          "family": "Sze"
        },
        {
          "given": "Y.-H.",
          "family": "Chen"
        },
        {
          "given": "T.-J.",
          "family": "Yang"
        },
        {
          "given": "J. S.",
          "family": "Emer"
        }
      ],
      "year": "2020",
      "title": "How to evaluate deep neural network processors: TOPS/W (Alone) considered harmful",
      "container": "IEEE Solid-State Circuits Magazine",
      "volume": "12",
      "issue": "3",
      "pages": "28-41",
      "doi": null,
      "url": null,
      "type": "journal-article"
    },
    {
      "raw_index": 3,
      "authors": [
        {
          "given": "A.",
          "family": "Moss"
        },
        {
          "given": "H.",
          "family": "Lee"
        },
        {
          "given": "L.",
          "family": "Xun"
        },
        {
          "given": "C.",
          "family": "Min"
        },
        {
          "given": "F.",
          "family": "Kawsar"
        },
        {
          "given": "A.",
          "family": "Montanari"
        }
      ],
      "year": "2022",
      "title": "Ultra-low-power DNN accelerators for IoT: Resource characterization of the MAX78000",
      "container": "SenSys",
      "volume": null,
      "issue": null,
      "pages": "934-940",
      "doi": null,
      "url": null,
      "type": "other"
    },
    {
      "raw_index": 4,
      "authors": [
        {
          "given": "M.",
          "family": "Clay"
        },
        {
          "given": "C.",
          "family": "Grecos"
        },
        {
          "given": "M.",
          "family": "Shirvaikar"
        },
        {
          "given": "B.",
          "family": "Richey"
        }
      ],
      "year": "2022",
      "title": "Benchmarking the MAX78000 artificial intelligence microcontroller for deep learning applications",
      "container": "Real-Time Image Processing and Deep Learning 2022, SPIE",
      "volume": null,
      "issue": null,
      "pages": null,
      "doi": null,
      "url": null,
      "type": "other"
    },
    {
      "raw_index": 5,
      "authors": [
        {
          "given": "Y.",
          "family": "Huang"
        },
        {
          "given": "T.",
          "family": "Gong"
        },
        {
          "given": "S.",
          "family": "Jang"
        },
        {
          "given": "F.",
          "family": "Kawsar"
        },
        {
          "given": "C.",
          "family": "Min"
        }
      ],
      "year": "2024",
      "title": "Energy characterization of tiny AI accelerator-equipped microcontrollers",
      "container": "HumanSys",
      "volume": null,
      "issue": null,
      "pages": "1-6",
      "doi": null,
      "url": null,
      "type": "other"
    },
    {
      "raw_index": 6,
      "authors": [
        {
          "given": "C.",
          "family": "Jeon"
        },
        {
          "given": "T.",
          "family": "Gong"
        },
        {
          "given": "J.",
          "family": "Yi"
        },
        {
          "given": "F.",
          "family": "Kawar"
        },
        {
          "given": "C.",
          "family": "Min"
        }
      ],
      "year": "2025",
      "title": "TinyMem: Boosting multi-DNN inference on tiny AI accelerators with weight memory virtualization",
      "container": "HotMobile",
      "volume": null,
      "issue": null,
      "pages": "1-6",
      "doi": null,
      "url": null,
      "type": "other"
    },
    {
      "raw_index": 7,
      "authors": [
        {
          "given": "T.",
          "family": "Gong"
        },
        {
          "given": "F.",
          "family": "Kawsar"
        },
        {
          "given": "C.",
          "family": "Min"
        }
      ],
      "year": "2025",
      "title": "DEX: Data channel extension for efficient CNN inference on tiny AI accelerators",
      "container": "NeurIPS",
      "volume": "37",
      "issue": null,
      "pages": "43925-43951",
      "doi": null,
      "url": null,
      "type": "other"
    },
    {
      "raw_index": 8,
      "authors": [
        {
          "given": null,
          "family": "Arm Ltd."
        }
      ],
      "year": "2024",
      "title": "Ethos-U Vela Compiler",
      "container": "Arm Developer Documentation",
      "volume": null,
      "issue": null,
      "pages": null,
      "doi": null,
      "url": "https://developer.arm.com/documentation/109267/0103/Tool-support-for-the-Arm-Ethos-U-NPU/Ethos-U-Vela-compiler",
      "type": "other"
    },
    {
      "raw_index": 9,
      "authors": [
        {
          "given": null,
          "family": "Analog Devices"
        }
      ],
      "year": "2024",
      "title": "MAX78000 Artificial Intelligence Microcontroller with Ultra-Low-Power Convolutional Neural Network Accelerator",
      "container": "Datasheet",
      "volume": "5",
      "issue": null,
      "pages": null,
      "doi": null,
      "url": "https://www.analog.com/media/en/technical-documentation/data-sheets/max78000.pdf",
      "type": "other"
    },
    {
      "raw_index": 10,
      "authors": [
        {
          "given": null,
          "family": "GreenWaves Technologies"
        }
      ],
      "year": "2020",
      "title": "GAP8 IoT Application Processor",
      "container": "Product Brief",
      "volume": "1",
      "issue": "9",
      "pages": null,
      "doi": null,
      "url": null,
      "type": "other"
    },
    {
      "raw_index": 11,
      "authors": [
        {
          "given": null,
          "family": "Himax Technologies"
        }
      ],
      "year": "2023",
      "title": "HX6538-A WE2 AI Processor",
      "container": "Preliminary Datasheet",
      "volume": null,
      "issue": null,
      "pages": null,
      "doi": null,
      "url": "https://files.seedstudio.com/wiki/grove-vision-ai-v2/HX6538_datasheet.pdf",
      "type": "other"
    },
    {
      "raw_index": 12,
      "authors": [
        {
          "given": null,
          "family": "NXP Semiconductors"
        }
      ],
      "year": "2024",
      "title": "FRDM-MCXN947 Board User Manual",
      "container": "User Manual UM12018",
      "volume": null,
      "issue": null,
      "pages": null,
      "doi": null,
      "url": "https://docs.nxp.com/bundle/UM12018/page/topics/Evk_overview.html",
      "type": "other"
    },
    {
      "raw_index": 13,
      "authors": [
        {
          "given": "S.",
          "family": "Laskaridis"
        },
        {
          "given": "A.",
          "family": "Kouris"
        },
        {
          "given": "N.D.",
          "family": "Lane"
        }
      ],
      "year": "2021",
      "title": "Adaptive inference through early-exit networks: Design, challenges and directions",
      "container": "Éditions Maison des Langues: EMDL",
      "volume": null,
      "issue": null,
      "pages": "1-6",
      "doi": null,
      "url": null,
      "type": "other"
    },
    {
      "raw_index": 14,
      "authors": [
        {
          "given": "V.J.",
          "family": "Reddi"
        },
        {
          "given": "D.",
          "family": "Kanter"
        },
        {
          "given": "P.",
          "family": "Mattson"
        },
        {
          "given": "J.",
          "family": "Duke"
        },
        {
          "given": "T.",
          "family": "Nguyen"
        },
        {
          "given": "R.",
          "family": "Chukka"
        },
        {
          "given": "K.",
          "family": "Shiring"
        },
        {
          "given": "K.-S.",
          "family": "Tan"
        },
        {
          "given": "M.",
          "family": "Charlebois"
        },
        {
          "given": "W.",
          "family": "Chou"
        }
      ],
      "year": "2022",
      "title": "MLPerf mobile inference benchmark: An industrystandard open-source machine learning benchmark for on-device AI",
      "container": "Proceedings of Machine Learning and Systems",
      "volume": "4",
      "issue": null,
      "pages": "352-369",
      "doi": null,
      "url": null,
      "type": "journal-article"
    }
  ],
  "_raw_references": [
    "S.S. Saha, S.S. Sandha, and M. Srivastava. 2022. Machine learning for microcontroller-class hardware: A review. _IEEE Sensors Journal_, 22(22):21362–21390.",
    "J. Lin, W.-M. Chen, Y. Lin, C. Gan, S. Han, et al. 2020. MCUNet: Tiny deep learning on IoT devices. _NeurIPS_, 33:11711–11722.",
    "V. Sze, Y.-H. Chen, T.-J. Yang, and J. S. Emer. 2020. How to evaluate deep neural network processors: TOPS/W (Alone) considered harmful. _IEEE Solid-State Circuits Magazine_, 12(3):28–41.",
    "A. Moss, H. Lee, L. Xun, C. Min, F. Kawsar, and A. Montanari. 2022. Ultra-low-power DNN accelerators for IoT: Resource characterization of the MAX78000. 2022. _SenSys_, 934–940.",
    "M. Clay, C. Grecos, M. Shirvaikar, and B. Richey. 2022. Benchmarking the MAX78000 artificial intelligence microcontroller for deep learning applications. _Real-Time Image Processing and Deep Learning_ 2022, SPIE.",
    "Y. Huang, T. Gong, S. Jang, F. Kawsar, and C. Min. 2024. Energy characterization of tiny AI accelerator-equipped microcontrollers. 2024. _HumanSys_, 1–6.",
    "C. Jeon, T. Gong, J. Yi, F. Kawar, and C. Min. 2025. TinyMem: Boosting multi-DNN inference on tiny AI accelerators with weight memory virtualization. 2025. _HotMobile_, 1-6.",
    "T. Gong, F. Kawsar, and C. Min. 2025. DEX: Data channel extension for efficient CNN inference on tiny AI accelerators. 2025. _NeurIPS_, 37:43925–43951.",
    "Arm Ltd. “Ethos-U Vela Compiler.” Arm Developer Documentation, Version 1.3, 2024. https://developer.arm.com/documentation/109267/0103/Tool-support-for-the-Arm-Ethos-U-NPU/Ethos-U-Vela-compiler",
    "Analog Devices. “MAX78000 Artificial Intelligence Microcontroller with Ultra-Low-Power Convolutional Neural Network Accelerator.” Datasheet, Rev. 5, 2024. https://www.analog.com/ media/en/technical-documentation/data-sheets/ max78000.pdf",
    "GreenWaves Technologies. “GAP8 IoT Applica- tion Processor.” Product Brief, Version 1.9, 2020.",
    "Himax Technologies. “HX6538-A WE2 AI Processor.” Preliminary Datasheet, Version 01, Jul. 2023. https://files.seedstudio.com/wiki/grove-vision-ai-v2/HX6538_datasheet.pdf",
    "NXP Semiconductors. “FRDM-MCXN947 Board User Manual.” User Manual UM12018, Rev. 2.0, Aug. 2024. https://docs.nxp.com/bundle/UM12018/page/topics/Evk_overview.html",
    "S. Laskaridis, A. Kouris, and N.D. Lane. 2021. Adaptive inference through early-exit networks: Design, challenges and directions.\" _Éditions Maison des Langues: EMDL_, 1-6.",
    "V.J. Reddi, D. Kanter, P. Mattson, J. Duke, T. Nguyen, R. Chukka, K. Shiring, K.-S. Tan, M. Charlebois, W. Chou, et al. 2022. MLPerf mobile inference benchmark: An industry-standard open-source machine learning benchmark for on-device AI. _Proceedings of Machine Learning and Systems_, 4:352–369."
  ],
  "_merge": {
    "grobid_tei": "/tmp/claude-1002/-home-avsm2-src-git-avsm-oxmono/e18896b5-df07-48cc-99c5-6308ff119d17/scratchpad/rerun/2026-benchmarking-npus/2026-benchmarking-npus.tei.xml",
    "header_fields_from_grobid": [],
    "resp": {
      "header": {},
      "references": {
        "b0": {
          "pages": "grobid",
          "title": "grobid"
        },
        "b1": {
          "pages": "grobid"
        },
        "b2": {
          "pages": "grobid"
        },
        "b3": {
          "pages": "grobid"
        },
        "b5": {
          "pages": "grobid"
        },
        "b7": {
          "pages": "grobid"
        },
        "b9": {
          "volume": "grobid"
        },
        "b10": {
          "issue": "grobid",
          "volume": "grobid"
        },
        "b14": {
          "pages": "grobid",
          "title": "grobid"
        }
      }
    }
  }
}