{"id":"https://openalex.org/W7224838560","doi":"https://doi.org/10.48550/arxiv.2610.40039","title":"How Far Are We From On-Device VLM Inference? A Measurement Study on Edge NPUs","display_name":"How Far Are We From On-Device VLM Inference? A Measurement Study on Edge NPUs","publication_year":2026,"publication_date":"2026-10-29","ids":{"openalex":"https://openalex.org/W7224838560","doi":"https://doi.org/10.48550/arxiv.2610.40039"},"language":"en","primary_location":{"id":"pmh:oai:arXiv.org:2610.40039","doi":null,"is_oa":true,"landing_page_url":"https://arxiv.org/abs/2610.40039","pdf_url":"https://arxiv.org/pdf/2610.40039","source":{"id":"https://openalex.org/S4306400194","display_name":"arXiv (Cornell University)","issn_l":"2331-8422","issn":["2331-8422"],"is_oa":true,"is_in_doaj":false,"is_core":false,"listed_in":[],"host_organization":"https://openalex.org/I205783295","host_organization_name":"Cornell University","host_organization_lineage":["https://openalex.org/I205783295"],"host_organization_lineage_names":[],"type":"repository"},"license":null,"license_id":null,"version":"submittedVersion","is_accepted":false,"is_published":false,"raw_source_name":null,"raw_type":"text"},"volume":null,"type":"preprint","indexed_in":["arxiv","datacite"],"open_access":{"is_oa":true,"oa_status":"green","oa_url":"https://arxiv.org/pdf/2610.40039","any_repository_has_fulltext":true},"authorships":[{"author_position":"first","author":{"id":"https://openalex.org/A5160173587","display_name":"Huan Yang","orcid":null,"observed_orcids":[]},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Yang, Huan","raw_affiliation_strings":["YU MAO ,"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"YU MAO ,","institution_ids":[]}]},{"author_position":"middle","author":{"id":"https://openalex.org/A5160216613","display_name":"Shiqi Jiang","orcid":null,"observed_orcids":[]},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Jiang, Shiqi","raw_affiliation_strings":["YU MAO ,"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"YU MAO ,","institution_ids":[]}]},{"author_position":"middle","author":{"id":"https://openalex.org/A5160183115","display_name":"Yu Mao","orcid":null,"observed_orcids":[]},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Mao, Yu","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5111911837","display_name":"Chuanyue Xiong","orcid":"https://orcid.org/0000-0002-9179-3530","observed_orcids":["https://orcid.org/0000-0002-9179-3530"]},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Xiong, Chuanyue","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"last","author":{"id":"https://openalex.org/A5160216819","display_name":"Deyu Zhang","orcid":null,"observed_orcids":[]},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Zhang, Deyu","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]}],"institutions":[],"countries_distinct_count":0,"institutions_distinct_count":0,"corresponding_author_ids":[],"corresponding_institution_ids":[],"apc_list":null,"apc_paid":null,"fwci":null,"has_fulltext":true,"cited_by_count":0,"citation_normalized_percentile":null,"cited_by_percentile_year":null,"biblio":{"volume":null,"issue":null,"first_page":null,"last_page":null},"is_retracted":false,"is_paratext":false,"is_xpac":false,"primary_topic":{"id":"https://openalex.org/T10054","display_name":"Parallel Computing and Optimization Techniques","score":0.4713999927043915,"subfield":{"id":"https://openalex.org/subfields/1708","display_name":"Hardware and Architecture"},"field":{"id":"https://openalex.org/fields/17","display_name":"Computer Science"},"domain":{"id":"https://openalex.org/domains/3","display_name":"Physical Sciences"}},"topics":[{"id":"https://openalex.org/T10054","display_name":"Parallel Computing and Optimization Techniques","score":0.4713999927043915,"subfield":{"id":"https://openalex.org/subfields/1708","display_name":"Hardware and Architecture"},"field":{"id":"https://openalex.org/fields/17","display_name":"Computer Science"},"domain":{"id":"https://openalex.org/domains/3","display_name":"Physical Sciences"}},{"id":"https://openalex.org/T10036","display_name":"Advanced Neural Network Applications","score":0.4214000105857849,"subfield":{"id":"https://openalex.org/subfields/1707","display_name":"Computer Vision and Pattern Recognition"},"field":{"id":"https://openalex.org/fields/17","display_name":"Computer Science"},"domain":{"id":"https://openalex.org/domains/3","display_name":"Physical Sciences"}},{"id":"https://openalex.org/T10273","display_name":"IoT and Edge/Fog Computing","score":0.061000000685453415,"subfield":{"id":"https://openalex.org/subfields/1705","display_name":"Computer Networks and Communications"},"field":{"id":"https://openalex.org/fields/17","display_name":"Computer Science"},"domain":{"id":"https://openalex.org/domains/3","display_name":"Physical Sciences"}}],"keywords":[{"id":"https://openalex.org/keywords/vision-language-models","display_name":"vision-language models","score":0.8679999709129333},{"id":"https://openalex.org/keywords/on-device-inference","display_name":"on-device inference","score":0.949999988079071},{"id":"https://openalex.org/keywords/linear-attention","display_name":"linear attention","score":0.7839999794960022},{"id":"https://openalex.org/keywords/autoregressive-decoding","display_name":"autoregressive decoding","score":0.8389999866485596},{"id":"https://openalex.org/keywords/pipeline-bubbles","display_name":"pipeline bubbles","score":0.5920000076293945},{"id":"https://openalex.org/keywords/inference-latency","display_name":"inference latency","score":0.5410000085830688},{"id":"https://openalex.org/keywords/cross-layer-optimization","display_name":"cross-layer optimization","score":0.625}],"concepts":[{"id":"https://openalex.org/C41008148","wikidata":"https://www.wikidata.org/wiki/Q21198","display_name":"Computer science","level":0,"score":0.6779999732971191},{"id":"https://openalex.org/C188441871","wikidata":"https://www.wikidata.org/wiki/Q7554146","display_name":"Softmax function","level":3,"score":0.6306999921798706},{"id":"https://openalex.org/C162307627","wikidata":"https://www.wikidata.org/wiki/Q204833","display_name":"Enhanced Data Rates for GSM Evolution","level":2,"score":0.5978999733924866},{"id":"https://openalex.org/C82876162","wikidata":"https://www.wikidata.org/wiki/Q17096504","display_name":"Latency (audio)","level":2,"score":0.5616999864578247},{"id":"https://openalex.org/C105339364","wikidata":"https://www.wikidata.org/wiki/Q2297740","display_name":"Software deployment","level":2,"score":0.557699978351593},{"id":"https://openalex.org/C43521106","wikidata":"https://www.wikidata.org/wiki/Q2165493","display_name":"Pipeline (software)","level":2,"score":0.4851999878883362},{"id":"https://openalex.org/C2776214188","wikidata":"https://www.wikidata.org/wiki/Q408386","display_name":"Inference","level":2,"score":0.4235999882221222},{"id":"https://openalex.org/C13164978","wikidata":"https://www.wikidata.org/wiki/Q600158","display_name":"Hardware acceleration","level":3,"score":0.4034999907016754},{"id":"https://openalex.org/C175309249","wikidata":"https://www.wikidata.org/wiki/Q725864","display_name":"Pipeline transport","level":2,"score":0.38260000944137573},{"id":"https://openalex.org/C173608175","wikidata":"https://www.wikidata.org/wiki/Q232661","display_name":"Parallel computing","level":1,"score":0.353300005197525},{"id":"https://openalex.org/C45374587","wikidata":"https://www.wikidata.org/wiki/Q12525525","display_name":"Computation","level":2,"score":0.3361999988555908},{"id":"https://openalex.org/C117896860","wikidata":"https://www.wikidata.org/wiki/Q11376","display_name":"Acceleration","level":2,"score":0.3192000091075897},{"id":"https://openalex.org/C2778456923","wikidata":"https://www.wikidata.org/wiki/Q5337692","display_name":"Edge computing","level":3,"score":0.3086000084877014},{"id":"https://openalex.org/C138236772","wikidata":"https://www.wikidata.org/wiki/Q25098575","display_name":"Edge device","level":3,"score":0.30730000138282776},{"id":"https://openalex.org/C159877910","wikidata":"https://www.wikidata.org/wiki/Q2202883","display_name":"Autoregressive model","level":2,"score":0.2971000075340271},{"id":"https://openalex.org/C79403827","wikidata":"https://www.wikidata.org/wiki/Q3988","display_name":"Real-time computing","level":1,"score":0.28360000252723694},{"id":"https://openalex.org/C2989134064","wikidata":"https://www.wikidata.org/wiki/Q288510","display_name":"Execution time","level":2,"score":0.27379998564720154},{"id":"https://openalex.org/C36503486","wikidata":"https://www.wikidata.org/wiki/Q11235244","display_name":"Domain (mathematical analysis)","level":2,"score":0.26980000734329224},{"id":"https://openalex.org/C180591934","wikidata":"https://www.wikidata.org/wiki/Q1253369","display_name":"Downtime","level":2,"score":0.2678000032901764}],"mesh":[],"locations_count":2,"locations":[{"id":"pmh:oai:arXiv.org:2610.40039","doi":null,"is_oa":true,"landing_page_url":"https://arxiv.org/abs/2610.40039","pdf_url":"https://arxiv.org/pdf/2610.40039","source":{"id":"https://openalex.org/S4306400194","display_name":"arXiv (Cornell University)","issn_l":"2331-8422","issn":["2331-8422"],"is_oa":true,"is_in_doaj":false,"is_core":false,"listed_in":[],"host_organization":"https://openalex.org/I205783295","host_organization_name":"Cornell University","host_organization_lineage":["https://openalex.org/I205783295"],"host_organization_lineage_names":[],"type":"repository"},"license":null,"license_id":null,"version":"submittedVersion","is_accepted":false,"is_published":false,"raw_source_name":null,"raw_type":"text"},{"id":"doi:10.48550/arxiv.2610.40039","doi":"https://doi.org/10.48550/arxiv.2610.40039","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2610.40039","pdf_url":null,"source":{"id":"https://openalex.org/S4306400194","display_name":"arXiv (Cornell University)","issn_l":"2331-8422","issn":["2331-8422"],"is_oa":true,"is_in_doaj":false,"is_core":false,"listed_in":[],"host_organization":"https://openalex.org/I205783295","host_organization_name":"Cornell University","host_organization_lineage":["https://openalex.org/I205783295"],"host_organization_lineage_names":[],"type":"repository"},"license":null,"license_id":null,"version":null,"is_accepted":false,"is_published":false,"raw_source_name":null,"raw_type":"Preprint"}],"best_oa_location":{"id":"pmh:oai:arXiv.org:2610.40039","doi":null,"is_oa":true,"landing_page_url":"https://arxiv.org/abs/2610.40039","pdf_url":"https://arxiv.org/pdf/2610.40039","source":{"id":"https://openalex.org/S4306400194","display_name":"arXiv (Cornell University)","issn_l":"2331-8422","issn":["2331-8422"],"is_oa":true,"is_in_doaj":false,"is_core":false,"listed_in":[],"host_organization":"https://openalex.org/I205783295","host_organization_name":"Cornell University","host_organization_lineage":["https://openalex.org/I205783295"],"host_organization_lineage_names":[],"type":"repository"},"license":null,"license_id":null,"version":"submittedVersion","is_accepted":false,"is_published":false,"raw_source_name":null,"raw_type":"text"},"sustainable_development_goals":[],"study_designs":[],"awards":[],"funders":[],"has_content":{"grobid_xml":true,"pdf":true},"content_urls":{"pdf":"https://content.openalex.org/works/W7224838560.pdf","grobid_xml":"https://content.openalex.org/works/W7224838560.grobid-xml"},"referenced_works_count":0,"referenced_works":[],"related_works":["https://openalex.org/W4417010825","https://openalex.org/W7202205216","https://openalex.org/W7163205670","https://openalex.org/W7214740290","https://openalex.org/W7164126783","https://openalex.org/W7139925212","https://openalex.org/W7214093136","https://openalex.org/W7168135622","https://openalex.org/W7164123379","https://openalex.org/W7162645558"],"abstract_inverted_index":{"Diverse":[0],"optimizations":[1,84],"spanning":[2],"model":[3,16,161],"architecture,":[4],"inference":[5],"runtime,":[6],"and":[7,49,61,123,160],"hardware":[8,112,156],"acceleration":[9],"have":[10],"been":[11],"proposed":[12],"for":[13,135,155],"on-device":[14],"vision-language":[15],"(VLM)":[17],"deployment":[18],"on":[19,28],"edge":[20,29,47],"NPUs.":[21],"Still,":[22],"cross-layer":[23,153],"coordination":[24],"of":[25],"VLM":[26],"execution":[27,68,104],"NPUs":[30],"remains":[31],"largely":[32],"underexplored.":[33],"This":[34],"paper":[35],"presents":[36],"an":[37],"empirical":[38],"measurement":[39],"study":[40],"across":[41],"six":[42],"representative":[43],"VLMs,":[44,136],"three":[45],"real-world":[46],"platforms,":[48],"ten":[50],"multimodal":[51],"benchmarks.":[52],"Our":[53],"evaluation":[54],"drills":[55],"down":[56,64],"from":[57,98],"end-to-end":[58],"application-level":[59],"latency":[60,145],"thermal":[62],"traces":[63],"to":[65,106],"low-level":[66],"kernel-scale":[67],"behaviors,":[69],"challenging":[70],"several":[71],"widely":[72],"held":[73],"community":[74],"assumptions.":[75],"First,":[76],"softmax":[77,83],"dominates":[78,115],"attention":[79,100],"costs;":[80],"even":[81],"lookup-table-based":[82],"cannot":[85],"eliminate":[86],"pipeline":[87],"bubbles":[88],"that":[89],"severely":[90],"underutilize":[91],"NPU":[92,121],"matrix":[93],"units.":[94],"Second,":[95],"theoretical":[96],"speedups":[97],"linear":[99],"vanish":[101],"under":[102],"real-NPU":[103],"due":[105],"prohibitive":[107],"tensor-rearrangement":[108],"overhead.":[109,146],"Third,":[110],"no":[111],"backend":[113],"universally":[114],"autoregressive":[116],"decoding:":[117],"attention-tensor-shaped":[118],"parallelism":[119],"dictates":[120],"efficiency":[122],"shifts":[124],"the":[125],"CPU-NPU":[126],"performance":[127],"balance.":[128],"We":[129],"further":[130],"find":[131],"task-specific":[132],"Sweet":[133],"Resolutions":[134],"which":[137],"deliver":[138],"meaningful":[139],"accuracy":[140],"gains":[141],"at":[142],"only":[143],"modest":[144],"Complementing":[147],"these":[148],"findings,":[149],"we":[150],"provide":[151],"actionable":[152],"recommendations":[154],"architects,":[157],"runtime":[158],"developers,":[159],"designers.":[162]},"counts_by_year":[],"updated_date":"2026-11-06T12:05:46.655746","created_date":"2026-10-31T00:00:00"}