[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"doc-detail-118435-en":3,"doc-seo-118435-105":30,"detail-sidebar-cat-0-en-105":91},{"code":4,"msg":5,"data":6},0,"success",{"doc_id":7,"user_id":8,"nickname":9,"user_avatar":10,"doc_module":4,"category_id":11,"category_name":12,"doc_title":13,"doc_description":14,"doc_content":15,"file_id":16,"file_url":17,"file_type":18,"file_size":19,"view_count":4,"is_deleted":4,"is_public":20,"is_downloadable":20,"audit_status":20,"page_count":21,"language":22,"language_code":23,"site_id":24,"html_lang":23,"table_of_contents":25,"faqs":26,"seo_title":27,"seo_description":14,"update_tm":28,"read_time":29},118435,687197207057,"Sage","https://ap-avatar.wpscdn.com/davatar_29158cc5080c5b710cf443261637dec0",8,"Research & Report","Analysis of Sub-Milliwatt Edge AI Accelerators for Modern Machine Learning Applications","Machine learning use continues to expand across industries, yet costly hardware and high power consumption limit adoption in small, sub-milliwatt deployments such as wireless sensor nodes. This thesis investigates whether compact sub-milliwatt edge AI accelerators can run common machine learning models effectively at substantially lower cost and power. The study outlines hardware selection and quantization-aware training, then evaluates experimentally collected benchmarks using metrics such as GOPS, GOPS per watt, energy per GOPS, and inference latency.","UC Irvine  \nUC Irvine Electronic Theses and Dissertations  \nTitle  \nAnalysis of Sub-Milliwatt Edge AI Accelerators for Modern Machine Learning Applications  \nPermalink  \n[https://escholarship.org/uc/item/7hx1p2f5](https://escholarship.org/uc/item/7hx1p2f5)  \nAuthor  \nArulnathan, Sarnesh  \nPublication Date  \n2025  \nCopyright Information  \nThis work is made available under the terms of a Creative Commons Attribution License, available at [https://creativecommons.org/licenses/by/4.0/](https://creativecommons.org/licenses/by/4.0/)  \n[Peer reviewed|Thesis/dissertation](Peer reviewed|Thesis/dissertation)  \n[eScholarship.org](eScholarship.org) Powered by the California Digital Library  \nUniversity of California  \nUNIVERSITY OF CALIFORNIA,  \nIRVINE  \nAnalysis of Sub-Milliwatt Edge AI Accelerators for Modern Machine Learning Applications  \nTHESIS  \nsubmitted in partial satisfaction of the requirements  \nfor the degree of  \nMASTER OF SCIENCE  \nin Software Engineering  \nby  \nSarnesh Arulnathan  \nTHESIS Committee:  \nAssistant Professor Sang-Woo Jun, Co-Chair Assistant Professor Joshua Garcia, Co-Chair Associate Professor Hadar Ziv  \n© 2025 Sarnesh Arulnathan  \nTABLE OF CONTENTS  \nPage  \nLIST OF FIGURES iv  \nLIST OF TABLES v  \nACKNOWLEDGMENTS vi  \nABSTRACT OF THE THESIS vii  \n1 Introduction 1  \n2 Motivation & Background 4  \n2.1 What is a sub-milliwatt edge AI accelerator .................. 4  \n2.2 A Large Problem ................................. 5  \n2.3 A Sub-milliwatt Solution ............................. 6  \n3 Overview 7  \n3.1 Approach ..................................... 7  \n3.2 Hardware Selection ................................ 8  \n3.3 Specs of sub-milliwatt edge AI accelerator ................... 9  \n3.4 Specs of over-milliwatt edge AI accelerators .................. 9  \n3.5 Specs of over-milliwatt non-AI accelerator ................... 10  \n3.6 Quantiazation-Aware Training (QAT) ...................... 11  \n3.7 Metrics ....................................... 11  \n4 Experimentally Collected Data 13  \n4.1 Notice to Readers ................................. 13  \n4.2 Visualizations ................................... 13  \n4.2.1 Chip Specs: Manufacturer, Power Usage, Cost, and Size ....... 14  \n4.2.2 Benchmark Results ............................ 14  \n5 Analysis 17  \n5.1 Cost and Power .................................. 17  \n5.2 Size Matters .................................... 18  \n5.3 GOPS ....................................... 18  \n5.4 GOPS per Watt .................................. 19  \n5.5 Energy per GOPS ................................. 20  \n5.6 Inference per Second ............................... 21  \n5.7 Cost per Inference ................................ 21  \n5.8 Limitations of sub-milliwatt AI accelerators .................. 22  \n5.9 Threats to Validity ................................ 23  \n6 Conclusion 24  \n6.0.1 Finding 1 ................................. 25  \n6.0.2 Finding 2 ................................. 25  \n6.0.3 Finding 3 ................................. 25  \n6.0.4 Finding 4 ................................. 25  \n6.1 Future Work .................................... 26  \nBibliography 27  \nLIST OF FIGURES  \nPage  \n2.1 The San Andreas Fault running through California’s major population centers. 5  \n3.1 A visual of the sub-milliwatt AI accelerator provided for this study...... 8  \n4.1 Comparison of Giga Operations Per Second (GOPS) Performances ...... 14  \n4.2 Comparison of Giga Operations Per Second (GOPS) per Wattage Consumed 15  \n4.3 Energy consumed in Joules over GOPS computed on the ResNet DNN ... 15  \n4.4 Inferences per Second computed on the MobileNet CNN ........... 16  \n4.5 Cost in USD per Inference computed per Second on a CNN model ...... 16  \nLIST OF TABLES  \nPage  \n4.1 Maximum Power Consumption, Cost, and Size of Chips ............ 14  \n5.1 Maximum Parameter Limitations for NDP 120 ................. 22  \nACKNOWLEDGMENTS  \nI thank Professor Joshua Garcia for being my committee chair and helping me place ","cbCaicq6NFDkuJxA","https://ap.wps.com/l/cbCaicq6NFDkuJxA","pdf",596103,1,37,"English","en",105,"# Introduction\n# Motivation & Background\n## What is a sub-milliwatt edge AI accelerator\n## A Large Problem\n## A Sub-milliwatt Solution\n# Overview\n## Approach\n## Hardware Selection\n## Quantiazation-Aware Training (QAT)\n## Metrics\n# Experimentally Collected Data\n## Visualizations\n# Analysis\n## Cost and Power\n## Size Matters\n## GOPS per Watt\n## Energy per GOPS\n## Inference per Second\n## Cost per Inference\n## Limitations of sub-milliwatt AI accelerators\n## Threats to Validity\n# Conclusion\n## Finding 1\n## Finding 2\n## Finding 3\n## Finding 4\n## Future Work","[{\"question\":\"Why does the thesis focus on sub-milliwatt edge AI accelerators?\",\"answer\":\"High hardware cost and power consumption hinder machine learning adoption in small deployments like wireless sensor nodes. Sub-milliwatt accelerators are explored as a way to enable more pervasive machine learning at lower cost and power.\"},{\"question\":\"What approach and hardware selection does the thesis use?\",\"answer\":\"The document presents an overview including an approach, hardware selection, and specification comparisons among sub-milliwatt and over-milliwatt edge AI accelerators as well as an over-milliwatt non-AI accelerator.\"},{\"question\":\"Which metrics are used to evaluate performance and efficiency?\",\"answer\":\"The analysis uses experimentally collected metrics including GOPS, GOPS per watt, energy per GOPS, inferences per second, and cost per inference, along with visualized benchmark results.\"}]","Analysis of Sub-Milliwatt Edge AI Accelerators for Modern Machine Learning Applications | PDF",1785683597,93,{"code":4,"msg":31,"data":32},"ok",{"site_id":24,"language":23,"slug":33,"title":13,"keywords":34,"description":14,"schema_data":35,"social_meta":86,"head_meta":88,"extra_data":90,"updated_unix":28},"analysis-of-sub-milliwatt-edge-ai-accelerators-for-modern-machine-learning-applications","",{"@graph":36,"@context":85},[37,54,68],{"@type":38,"itemListElement":39},"BreadcrumbList",[40,44,48,51],{"item":41,"name":42,"@type":43,"position":20},"https://docshare.wps.com","Home","ListItem",{"item":45,"name":46,"@type":43,"position":47},"https://docshare.wps.com/document/","Document",2,{"item":49,"name":12,"@type":43,"position":50},"https://docshare.wps.com/document/research-report/",3,{"item":52,"name":13,"@type":43,"position":53},"https://docshare.wps.com/document/analysis-of-sub-milliwatt-edge-ai-accelerators-for-modern-machine-learning-applications/118435/",4,{"url":52,"name":13,"@type":55,"author":56,"headline":13,"publisher":58,"fileFormat":61,"inLanguage":23,"description":14,"dateModified":62,"datePublished":62,"encodingFormat":61,"isAccessibleForFree":63,"interactionStatistic":64},"DigitalDocument",{"name":9,"@type":57},"Person",{"url":41,"name":59,"@type":60},"DocShare","Organization","application/pdf","2026-08-02",true,{"@type":65,"interactionType":66,"userInteractionCount":4},"InteractionCounter",{"@type":67},"ViewAction",{"@type":69,"mainEntity":70},"FAQPage",[71,77,81],{"name":72,"@type":73,"acceptedAnswer":74},"Why does the thesis focus on sub-milliwatt edge AI accelerators?","Question",{"text":75,"@type":76},"High hardware cost and power consumption hinder machine learning adoption in small deployments like wireless sensor nodes. Sub-milliwatt accelerators are explored as a way to enable more pervasive machine learning at lower cost and power.","Answer",{"name":78,"@type":73,"acceptedAnswer":79},"What approach and hardware selection does the thesis use?",{"text":80,"@type":76},"The document presents an overview including an approach, hardware selection, and specification comparisons among sub-milliwatt and over-milliwatt edge AI accelerators as well as an over-milliwatt non-AI accelerator.",{"name":82,"@type":73,"acceptedAnswer":83},"Which metrics are used to evaluate performance and efficiency?",{"text":84,"@type":76},"The analysis uses experimentally collected metrics including GOPS, GOPS per watt, energy per GOPS, inferences per second, and cost per inference, along with visualized benchmark results.","https://schema.org",{"og:url":52,"og:type":87,"og:title":13,"og:site_name":59,"og:description":14},"article",{"robots":89,"canonical":52},"index,follow",{"doc_id":7,"site_id":24},{"code":4,"msg":5,"data":92},[93,97,101,105,110,115,120,123,128,131,135],{"id":20,"doc_module":4,"doc_module_name":46,"category_name":94,"show_sort_weight":95,"slug":96},"Story & Novel",90,"story-novel",{"id":47,"doc_module":4,"doc_module_name":46,"category_name":98,"show_sort_weight":99,"slug":100},"Literature",80,"literature",{"id":53,"doc_module":4,"doc_module_name":46,"category_name":102,"show_sort_weight":103,"slug":104},"Exam",70,"exam",{"id":106,"doc_module":4,"doc_module_name":46,"category_name":107,"show_sort_weight":108,"slug":109},5,"Comic",60,"comic",{"id":111,"doc_module":4,"doc_module_name":46,"category_name":112,"show_sort_weight":113,"slug":114},6,"Technology",50,"technology",{"id":116,"doc_module":4,"doc_module_name":46,"category_name":117,"show_sort_weight":118,"slug":119},7,"Healthcare",40,"healthcare",{"id":11,"doc_module":4,"doc_module_name":46,"category_name":12,"show_sort_weight":121,"slug":122},30,"research-report",{"id":124,"doc_module":4,"doc_module_name":46,"category_name":125,"show_sort_weight":126,"slug":127},9,"Religion & Spirituality",20,"religion-spirituality",{"id":126,"doc_module":4,"doc_module_name":46,"category_name":129,"show_sort_weight":126,"slug":130},"World Cup","world-cup",{"id":132,"doc_module":4,"doc_module_name":46,"category_name":133,"show_sort_weight":132,"slug":134},10,"Lifestyle","lifestyle",{"id":136,"doc_module":4,"doc_module_name":46,"category_name":137,"show_sort_weight":106,"slug":138},19,"General","general"]