{"id":"https://openalex.org/W7128375882","doi":"https://doi.org/10.48550/arxiv.2602.06063","title":"Mapping Gemma3 onto an Edge Dataflow Architecture","display_name":"Mapping Gemma3 onto an Edge Dataflow Architecture","publication_year":2026,"publication_date":"2026-01-27","ids":{"openalex":"https://openalex.org/W7128375882","doi":"https://doi.org/10.48550/arxiv.2602.06063"},"language":"en","primary_location":{"id":"pmh:oai:arXiv.org:2602.06063","doi":null,"is_oa":true,"landing_page_url":"http://arxiv.org/abs/2602.06063","pdf_url":"https://arxiv.org/pdf/2602.06063","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/2602.06063","any_repository_has_fulltext":true},"authorships":[{"author_position":"first","author":{"id":"https://openalex.org/A5025040115","display_name":"Shouyu Du","orcid":"https://orcid.org/0009-0000-8321-0457","observed_orcids":["https://orcid.org/0009-0000-8321-0457"]},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Du, Shouyu","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5013631588","display_name":"Miaoxiang Yu","orcid":"https://orcid.org/0000-0002-4382-9009","observed_orcids":["https://orcid.org/0000-0002-4382-9009"]},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Yu, Miaoxiang","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5142603810","display_name":"Zhenyu Xu","orcid":null,"observed_orcids":[]},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Xu, Zhenyu","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5138299242","display_name":"Zhiheng Ni","orcid":null,"observed_orcids":[]},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Ni, Zhiheng","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5020091845","display_name":"Jillian Cai","orcid":"https://orcid.org/0009-0001-5006-223X","observed_orcids":["https://orcid.org/0009-0001-5006-223X"]},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Cai, Jillian","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5134844704","display_name":"Qing Yang","orcid":null,"observed_orcids":[]},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Yang, Qing","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"last","author":{"id":"https://openalex.org/A5137544062","display_name":"Tao Wei","orcid":null,"observed_orcids":[]},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Wei, Tao","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":false,"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":0.70169997215271,"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":0.70169997215271,"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":0.23729999363422394,"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/T10904","display_name":"Embedded Systems Design Techniques","score":0.016300000250339508,"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"}}],"keywords":[{"id":"https://openalex.org/keywords/gemma-3","display_name":"Gemma 3","score":0.984000027179718},{"id":"https://openalex.org/keywords/large-language-models","display_name":"large language models","score":0.8320000171661377},{"id":"https://openalex.org/keywords/vision-language-models","display_name":"vision-language models","score":0.9319999814033508},{"id":"https://openalex.org/keywords/edge-inference","display_name":"edge inference","score":0.6949999928474426},{"id":"https://openalex.org/keywords/dequantization","display_name":"dequantization","score":0.6570000052452087},{"id":"https://openalex.org/keywords/attention-mechanism","display_name":"attention mechanism","score":0.5360000133514404},{"id":"https://openalex.org/keywords/memory-bandwidth","display_name":"memory bandwidth","score":0.593999981880188},{"id":"https://openalex.org/keywords/power-efficiency","display_name":"power efficiency","score":0.6110000014305115}],"concepts":[{"id":"https://openalex.org/C96324660","wikidata":"https://www.wikidata.org/wiki/Q205446","display_name":"Dataflow","level":2,"score":0.8334000110626221},{"id":"https://openalex.org/C41008148","wikidata":"https://www.wikidata.org/wiki/Q21198","display_name":"Computer science","level":0,"score":0.7975999712944031},{"id":"https://openalex.org/C176727019","wikidata":"https://www.wikidata.org/wiki/Q1172415","display_name":"Dataflow architecture","level":3,"score":0.534600019454956},{"id":"https://openalex.org/C173608175","wikidata":"https://www.wikidata.org/wiki/Q232661","display_name":"Parallel computing","level":1,"score":0.49480000138282776},{"id":"https://openalex.org/C118524514","wikidata":"https://www.wikidata.org/wiki/Q173212","display_name":"Computer architecture","level":1,"score":0.44859999418258667},{"id":"https://openalex.org/C28855332","wikidata":"https://www.wikidata.org/wiki/Q198099","display_name":"Quantization (signal processing)","level":2,"score":0.4343999922275543},{"id":"https://openalex.org/C2776214188","wikidata":"https://www.wikidata.org/wiki/Q408386","display_name":"Inference","level":2,"score":0.4097000062465668},{"id":"https://openalex.org/C57273362","wikidata":"https://www.wikidata.org/wiki/Q576722","display_name":"Decoding methods","level":2,"score":0.399399995803833},{"id":"https://openalex.org/C123657996","wikidata":"https://www.wikidata.org/wiki/Q12271","display_name":"Architecture","level":2,"score":0.397599995136261},{"id":"https://openalex.org/C105339364","wikidata":"https://www.wikidata.org/wiki/Q2297740","display_name":"Software deployment","level":2,"score":0.3874000012874603},{"id":"https://openalex.org/C14580979","wikidata":"https://www.wikidata.org/wiki/Q876049","display_name":"Very-large-scale integration","level":2,"score":0.3817000091075897},{"id":"https://openalex.org/C138236772","wikidata":"https://www.wikidata.org/wiki/Q25098575","display_name":"Edge device","level":3,"score":0.3774000108242035},{"id":"https://openalex.org/C177264268","wikidata":"https://www.wikidata.org/wiki/Q1514741","display_name":"Set (abstract data type)","level":2,"score":0.36739999055862427},{"id":"https://openalex.org/C17349429","wikidata":"https://www.wikidata.org/wiki/Q1049914","display_name":"Matrix multiplication","level":3,"score":0.35659998655319214},{"id":"https://openalex.org/C9390403","wikidata":"https://www.wikidata.org/wiki/Q3966","display_name":"Computer hardware","level":1,"score":0.35409998893737793},{"id":"https://openalex.org/C162307627","wikidata":"https://www.wikidata.org/wiki/Q204833","display_name":"Enhanced Data Rates for GSM Evolution","level":2,"score":0.337799996137619},{"id":"https://openalex.org/C155911762","wikidata":"https://www.wikidata.org/wiki/Q422321","display_name":"Blueprint","level":2,"score":0.3301999866962433},{"id":"https://openalex.org/C11413529","wikidata":"https://www.wikidata.org/wiki/Q8366","display_name":"Algorithm","level":1,"score":0.32670000195503235},{"id":"https://openalex.org/C149635348","wikidata":"https://www.wikidata.org/wiki/Q193040","display_name":"Embedded system","level":1,"score":0.3206000030040741},{"id":"https://openalex.org/C2780615836","wikidata":"https://www.wikidata.org/wiki/Q2471869","display_name":"USable","level":2,"score":0.3140000104904175},{"id":"https://openalex.org/C199360897","wikidata":"https://www.wikidata.org/wiki/Q9143","display_name":"Programming language","level":1,"score":0.3059999942779541},{"id":"https://openalex.org/C3115412","wikidata":"https://www.wikidata.org/wiki/Q1194708","display_name":"Bitmap","level":2,"score":0.2962999939918518},{"id":"https://openalex.org/C2776257435","wikidata":"https://www.wikidata.org/wiki/Q1576430","display_name":"Bandwidth (computing)","level":2,"score":0.296099990606308},{"id":"https://openalex.org/C80444323","wikidata":"https://www.wikidata.org/wiki/Q2878974","display_name":"Theoretical computer science","level":1,"score":0.2921000123023987},{"id":"https://openalex.org/C175291020","wikidata":"https://www.wikidata.org/wiki/Q1156822","display_name":"Offset (computer science)","level":2,"score":0.2890999913215637},{"id":"https://openalex.org/C113775141","wikidata":"https://www.wikidata.org/wiki/Q428691","display_name":"Computer engineering","level":1,"score":0.28540000319480896},{"id":"https://openalex.org/C106487976","wikidata":"https://www.wikidata.org/wiki/Q685816","display_name":"Matrix (chemical analysis)","level":2,"score":0.27630001306533813},{"id":"https://openalex.org/C188045654","wikidata":"https://www.wikidata.org/wiki/Q17148339","display_name":"Memory bandwidth","level":2,"score":0.27570000290870667},{"id":"https://openalex.org/C94375191","wikidata":"https://www.wikidata.org/wiki/Q11205","display_name":"Arithmetic","level":1,"score":0.27129998803138733},{"id":"https://openalex.org/C98025372","wikidata":"https://www.wikidata.org/wiki/Q477538","display_name":"Systems architecture","level":3,"score":0.26440000534057617},{"id":"https://openalex.org/C120314980","wikidata":"https://www.wikidata.org/wiki/Q180634","display_name":"Distributed computing","level":1,"score":0.2614000141620636},{"id":"https://openalex.org/C121684516","wikidata":"https://www.wikidata.org/wiki/Q7600677","display_name":"Computer graphics (images)","level":1,"score":0.257999986410141},{"id":"https://openalex.org/C154945302","wikidata":"https://www.wikidata.org/wiki/Q11660","display_name":"Artificial intelligence","level":1,"score":0.2547999918460846},{"id":"https://openalex.org/C2778456923","wikidata":"https://www.wikidata.org/wiki/Q5337692","display_name":"Edge computing","level":3,"score":0.25360000133514404},{"id":"https://openalex.org/C26517878","wikidata":"https://www.wikidata.org/wiki/Q228039","display_name":"Key (lock)","level":2,"score":0.25279998779296875},{"id":"https://openalex.org/C118993495","wikidata":"https://www.wikidata.org/wiki/Q5042828","display_name":"Electrical efficiency","level":3,"score":0.2517000138759613}],"mesh":[],"locations_count":2,"locations":[{"id":"pmh:oai:arXiv.org:2602.06063","doi":null,"is_oa":true,"landing_page_url":"http://arxiv.org/abs/2602.06063","pdf_url":"https://arxiv.org/pdf/2602.06063","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.2602.06063","doi":"https://doi.org/10.48550/arxiv.2602.06063","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2602.06063","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:2602.06063","doi":null,"is_oa":true,"landing_page_url":"http://arxiv.org/abs/2602.06063","pdf_url":"https://arxiv.org/pdf/2602.06063","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":false,"pdf":false},"content_urls":null,"referenced_works_count":0,"referenced_works":[],"related_works":["https://openalex.org/W7164320377","https://openalex.org/W7214740290","https://openalex.org/W7105184807","https://openalex.org/W7163527459","https://openalex.org/W7164149133","https://openalex.org/W6910635050","https://openalex.org/W4415307300","https://openalex.org/W4413442950","https://openalex.org/W7160191658","https://openalex.org/W4415065479"],"abstract_inverted_index":{"We":[0],"present":[1],"the":[2,7,101,108,123,143],"first":[3],"end-to-end":[4],"deployment":[5],"of":[6,10,31],"Gemma3":[8],"family":[9],"large":[11],"language":[12],"and":[13,47,63,69,96,103,105,119,125,139,145],"vision":[14],"models":[15,153],"on":[16],"a":[17,29,50,66,82,147],"tiled":[18,43,155],"edge":[19],"dataflow":[20,156],"architecture":[21],"(AMD":[22],"Ryzen":[23],"AI":[24],"NPU).":[25],"Our":[26],"work":[27],"introduces":[28],"set":[30],"hardware-aware":[32],"techniques.":[33],"For":[34,55],"prefill,":[35],"we":[36,57],"introduce":[37,58],"an":[38],"efficient":[39],"dequantization":[40,62],"engine,":[41],"optimize":[42],"matrix":[44],"multiplication":[45],"kernels,":[46],"propose":[48],"FlowQKV,":[49],"chunked,":[51],"pipelined":[52],"attention":[53,73],"mechanism.":[54],"decoding,":[56],"FusedDQP,":[59],"which":[60,71],"fuses":[61],"projection":[64],"into":[65],"single":[67],"kernel,":[68],"FlowKV,":[70],"re-structures":[72],"to":[74,92,122],"sustain":[75],"high":[76],"memory":[77],"bandwidth":[78],"utilization.":[79],"Together":[80],"with":[81],"compact":[83],"Q4NX":[84],"4-bit":[85],"quantization":[86],"format,":[87],"these":[88],"methods":[89],"yield":[90],"up":[91],"$5.2\\times$":[93],"faster":[94,98],"prefill":[95],"$4.8\\times$":[97],"decoding":[99],"versus":[100],"iGPU,":[102],"$33.5\\times$":[104],"$2.2\\times$":[106],"over":[107],"CPU,":[109],"respectively.":[110],"Power":[111],"efficiency":[112],"improves":[113],"by":[114],"as":[115,117],"much":[116],"$67.2\\times$":[118],"$222.9\\times$":[120],"compared":[121],"iGPU":[124],"CPU.":[126],"The":[127],"proposed":[128],"approach":[129],"demonstrates":[130],"that":[131],"modern":[132],"NPUs":[133],"can":[134],"deliver":[135],"practical,":[136],"low-power":[137],"LLM":[138],"VLM":[140],"inference":[141],"at":[142],"edge,":[144],"provides":[146],"generalizable":[148],"blueprint":[149],"for":[150],"mapping":[151],"transformer-based":[152],"onto":[154],"accelerators.":[157]},"counts_by_year":[],"updated_date":"2026-10-28T10:33:24.094632","created_date":"2026-02-10T00:00:00"}