{"id":"https://openalex.org/W4413442950","doi":"https://doi.org/10.1109/coins65080.2025.11125797","title":"LLM-NPU: Towards Efficient Foundation Model Inference on Low-Power Neural Processing Units","display_name":"LLM-NPU: Towards Efficient Foundation Model Inference on Low-Power Neural Processing Units","publication_year":2025,"publication_date":"2025-08-04","ids":{"openalex":"https://openalex.org/W4413442950","doi":"https://doi.org/10.1109/coins65080.2025.11125797"},"language":"en","primary_location":{"id":"doi:10.1109/coins65080.2025.11125797","is_oa":false,"landing_page_url":"https://doi.org/10.1109/coins65080.2025.11125797","pdf_url":null,"source":null,"license":null,"license_id":null,"version":"publishedVersion","is_accepted":true,"is_published":true,"raw_source_name":"2025 IEEE International Conference on Omni-layer Intelligent Systems (COINS)","raw_type":"proceedings-article"},"type":"conference-paper","indexed_in":["crossref"],"open_access":{"is_oa":false,"oa_status":"closed","oa_url":null,"any_repository_has_fulltext":false},"authorships":[{"author_position":"first","author":{"id":"https://openalex.org/A5066464351","display_name":"Arnab Raha","orcid":"https://orcid.org/0000-0002-8848-1069"},"institutions":[{"id":"https://openalex.org/I1343180700","display_name":"Intel (United States)","ror":"https://ror.org/01ek73717","country_code":"US","type":"company","lineage":["https://openalex.org/I1343180700"]}],"countries":["US"],"is_corresponding":false,"raw_author_name":"Arnab Raha","raw_affiliation_strings":["NPU IP, Intel Corporation,USA"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"NPU IP, Intel Corporation,USA","institution_ids":["https://openalex.org/I1343180700"]}]},{"author_position":"middle","author":{"id":"https://openalex.org/A5076002956","display_name":"Souvik Kundu","orcid":"https://orcid.org/0000-0002-3533-9405"},"institutions":[{"id":"https://openalex.org/I1343180700","display_name":"Intel (United States)","ror":"https://ror.org/01ek73717","country_code":"US","type":"company","lineage":["https://openalex.org/I1343180700"]}],"countries":["US"],"is_corresponding":false,"raw_author_name":"Souvik Kundu","raw_affiliation_strings":["Intel Labs,USA","NPU IP, Intel Corporation,USA"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"Intel Labs,USA","institution_ids":["https://openalex.org/I1343180700"]},{"raw_affiliation_string":"NPU IP, Intel Corporation,USA","institution_ids":["https://openalex.org/I1343180700"]}]},{"author_position":"middle","author":{"id":"https://openalex.org/A5101153316","display_name":"Sharath Nittur Sridhar","orcid":null},"institutions":[{"id":"https://openalex.org/I1343180700","display_name":"Intel (United States)","ror":"https://ror.org/01ek73717","country_code":"US","type":"company","lineage":["https://openalex.org/I1343180700"]}],"countries":["US"],"is_corresponding":false,"raw_author_name":"Sharath Nittur Sridhar","raw_affiliation_strings":["Intel Labs,USA"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"Intel Labs,USA","institution_ids":["https://openalex.org/I1343180700"]}]},{"author_position":"middle","author":{"id":"https://openalex.org/A5066094401","display_name":"Shamik Kundu","orcid":"https://orcid.org/0000-0002-5992-8554"},"institutions":[{"id":"https://openalex.org/I1343180700","display_name":"Intel (United States)","ror":"https://ror.org/01ek73717","country_code":"US","type":"company","lineage":["https://openalex.org/I1343180700"]}],"countries":["US"],"is_corresponding":false,"raw_author_name":"Shamik Kundu","raw_affiliation_strings":["Intel Labs,USA","NPU IP, Intel Corporation,USA"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"Intel Labs,USA","institution_ids":["https://openalex.org/I1343180700"]},{"raw_affiliation_string":"NPU IP, Intel Corporation,USA","institution_ids":["https://openalex.org/I1343180700"]}]},{"author_position":"middle","author":{"id":"https://openalex.org/A5043334127","display_name":"Soumendu Kumar Ghosh","orcid":"https://orcid.org/0000-0001-6776-1427"},"institutions":[{"id":"https://openalex.org/I1343180700","display_name":"Intel (United States)","ror":"https://ror.org/01ek73717","country_code":"US","type":"company","lineage":["https://openalex.org/I1343180700"]}],"countries":["US"],"is_corresponding":false,"raw_author_name":"Soumendu Kumar Ghosh","raw_affiliation_strings":["NPU IP, Intel Corporation,USA"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"NPU IP, Intel Corporation,USA","institution_ids":["https://openalex.org/I1343180700"]}]},{"author_position":"middle","author":{"id":"https://openalex.org/A5031637567","display_name":"Alessandro Palla","orcid":"https://orcid.org/0009-0005-9253-7191"},"institutions":[{"id":"https://openalex.org/I1343180700","display_name":"Intel (United States)","ror":"https://ror.org/01ek73717","country_code":"US","type":"company","lineage":["https://openalex.org/I1343180700"]}],"countries":["US"],"is_corresponding":false,"raw_author_name":"Alessandro Palla","raw_affiliation_strings":["NPU IP, Intel Corporation,USA"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"NPU IP, Intel Corporation,USA","institution_ids":["https://openalex.org/I1343180700"]}]},{"author_position":"middle","author":{"id":"https://openalex.org/A5102891312","display_name":"Arghadip Das","orcid":"https://orcid.org/0000-0001-6043-4685"},"institutions":[{"id":"https://openalex.org/I219193219","display_name":"Purdue University West Lafayette","ror":"https://ror.org/02dqehb95","country_code":"US","type":"education","lineage":["https://openalex.org/I219193219"]}],"countries":["US"],"is_corresponding":false,"raw_author_name":"Arghadip Das","raw_affiliation_strings":["Purdue University,USA"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"Purdue University,USA","institution_ids":["https://openalex.org/I219193219"]}]},{"author_position":"middle","author":{"id":"https://openalex.org/A5006678693","display_name":"Darren Crews","orcid":null},"institutions":[{"id":"https://openalex.org/I1343180700","display_name":"Intel (United States)","ror":"https://ror.org/01ek73717","country_code":"US","type":"company","lineage":["https://openalex.org/I1343180700"]}],"countries":["US"],"is_corresponding":false,"raw_author_name":"Darren Crews","raw_affiliation_strings":["NPU IP, Intel Corporation,USA"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"NPU IP, Intel Corporation,USA","institution_ids":["https://openalex.org/I1343180700"]}]},{"author_position":"last","author":{"id":"https://openalex.org/A5052021148","display_name":"Deepak A. Mathaikutty","orcid":"https://orcid.org/0009-0004-3768-1337"},"institutions":[{"id":"https://openalex.org/I1343180700","display_name":"Intel (United States)","ror":"https://ror.org/01ek73717","country_code":"US","type":"company","lineage":["https://openalex.org/I1343180700"]}],"countries":["US"],"is_corresponding":false,"raw_author_name":"Deepak A. Mathaikutty","raw_affiliation_strings":["NPU IP, Intel Corporation,USA"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"NPU IP, Intel Corporation,USA","institution_ids":["https://openalex.org/I1343180700"]}]}],"institutions":[],"countries_distinct_count":1,"institutions_distinct_count":2,"corresponding_author_ids":[],"corresponding_institution_ids":[],"apc_list":null,"apc_paid":null,"fwci":2.8273,"has_fulltext":false,"cited_by_count":2,"citation_normalized_percentile":{"value":0.91625227,"is_in_top_1_percent":false,"is_in_top_10_percent":true},"cited_by_percentile_year":{"min":95,"max":96},"biblio":{"volume":null,"issue":null,"first_page":"1","last_page":"8"},"is_retracted":false,"is_paratext":false,"is_xpac":false,"primary_topic":{"id":"https://openalex.org/T10036","display_name":"Advanced Neural Network Applications","score":0.998199999332428,"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"}},"topics":[{"id":"https://openalex.org/T10036","display_name":"Advanced Neural Network Applications","score":0.998199999332428,"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/T10320","display_name":"Neural Networks and Applications","score":0.9951000213623047,"subfield":{"id":"https://openalex.org/subfields/1702","display_name":"Artificial Intelligence"},"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/T10054","display_name":"Parallel Computing and Optimization Techniques","score":0.984499990940094,"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/foundation","display_name":"Foundation (evidence)","score":0.6526570320129395},{"id":"https://openalex.org/keywords/inference","display_name":"Inference","score":0.6402291059494019},{"id":"https://openalex.org/keywords/computer-science","display_name":"Computer science","score":0.6126910448074341},{"id":"https://openalex.org/keywords/artificial-neural-network","display_name":"Artificial neural network","score":0.5135695934295654},{"id":"https://openalex.org/keywords/power","display_name":"Power (physics)","score":0.48016729950904846},{"id":"https://openalex.org/keywords/artificial-intelligence","display_name":"Artificial intelligence","score":0.4448370039463043}],"concepts":[{"id":"https://openalex.org/C2780966255","wikidata":"https://www.wikidata.org/wiki/Q5474306","display_name":"Foundation (evidence)","level":2,"score":0.6526570320129395},{"id":"https://openalex.org/C2776214188","wikidata":"https://www.wikidata.org/wiki/Q408386","display_name":"Inference","level":2,"score":0.6402291059494019},{"id":"https://openalex.org/C41008148","wikidata":"https://www.wikidata.org/wiki/Q21198","display_name":"Computer science","level":0,"score":0.6126910448074341},{"id":"https://openalex.org/C50644808","wikidata":"https://www.wikidata.org/wiki/Q192776","display_name":"Artificial neural network","level":2,"score":0.5135695934295654},{"id":"https://openalex.org/C163258240","wikidata":"https://www.wikidata.org/wiki/Q25342","display_name":"Power (physics)","level":2,"score":0.48016729950904846},{"id":"https://openalex.org/C154945302","wikidata":"https://www.wikidata.org/wiki/Q11660","display_name":"Artificial intelligence","level":1,"score":0.4448370039463043},{"id":"https://openalex.org/C95457728","wikidata":"https://www.wikidata.org/wiki/Q309","display_name":"History","level":0,"score":0.0},{"id":"https://openalex.org/C121332964","wikidata":"https://www.wikidata.org/wiki/Q413","display_name":"Physics","level":0,"score":0.0},{"id":"https://openalex.org/C62520636","wikidata":"https://www.wikidata.org/wiki/Q944","display_name":"Quantum mechanics","level":1,"score":0.0},{"id":"https://openalex.org/C166957645","wikidata":"https://www.wikidata.org/wiki/Q23498","display_name":"Archaeology","level":1,"score":0.0}],"mesh":[],"locations_count":1,"locations":[{"id":"doi:10.1109/coins65080.2025.11125797","is_oa":false,"landing_page_url":"https://doi.org/10.1109/coins65080.2025.11125797","pdf_url":null,"source":null,"license":null,"license_id":null,"version":"publishedVersion","is_accepted":true,"is_published":true,"raw_source_name":"2025 IEEE International Conference on Omni-layer Intelligent Systems (COINS)","raw_type":"proceedings-article"}],"best_oa_location":null,"sustainable_development_goals":[],"awards":[],"funders":[],"has_content":{"pdf":false,"grobid_xml":false},"content_urls":null,"referenced_works_count":26,"referenced_works":["https://openalex.org/W2094756095","https://openalex.org/W2979826702","https://openalex.org/W2990591126","https://openalex.org/W3183327882","https://openalex.org/W3190761184","https://openalex.org/W3202917172","https://openalex.org/W4280590405","https://openalex.org/W4307934016","https://openalex.org/W4330336692","https://openalex.org/W4384918448","https://openalex.org/W4387199936","https://openalex.org/W4387725429","https://openalex.org/W4392739458","https://openalex.org/W4392746406","https://openalex.org/W4394778349","https://openalex.org/W4400489024","https://openalex.org/W4402671766","https://openalex.org/W4403364024","https://openalex.org/W4404261157","https://openalex.org/W4404570501","https://openalex.org/W4405469621","https://openalex.org/W4407196790","https://openalex.org/W4407358771","https://openalex.org/W4408182386","https://openalex.org/W4411664729","https://openalex.org/W4412889558"],"related_works":["https://openalex.org/W2381393187","https://openalex.org/W2332779545","https://openalex.org/W2358060160","https://openalex.org/W2035483685","https://openalex.org/W1969764885","https://openalex.org/W596947562","https://openalex.org/W2793937822","https://openalex.org/W2790817834","https://openalex.org/W2777605427","https://openalex.org/W2501983714"],"abstract_inverted_index":{"Neural":[0],"processing":[1],"units":[2],"(NPUs)":[3],"have":[4],"become":[5],"essential":[6],"in":[7,125],"modern":[8],"client":[9],"and":[10,69,92,102,115,122,155,168],"edge":[11],"platforms,":[12],"offering":[13],"unparalleled":[14],"efficiency":[15],"by":[16],"delivering":[17],"high":[18],"throughput":[19],"at":[20],"low":[21],"power.":[22],"This":[23,72],"is":[24],"critical":[25],"to":[26,34,43,56],"improve":[27,164],"the":[28,31,53,129,165],"TOPS/W":[29],"of":[30,170],"NPU,":[32],"leading":[33],"longer":[35],"battery":[36],"life.":[37],"While":[38],"NPUs":[39,88,171],"were":[40],"initially":[41],"designed":[42],"efficiently":[44],"execute":[45],"computer":[46],"vision":[47],"(CV)":[48],"workloads":[49],"such":[50,99],"as":[51,100],"CNNs,":[52],"rising":[54],"demand":[55],"run":[57],"transformer-based":[58],"large":[59],"language":[60],"models":[61],"(LLMs)":[62],"locally":[63],"now":[64],"calls":[65],"for":[66,172],"significant":[67],"architectural":[68],"software":[70,97],"adaptation.":[71],"paper":[73],"presents":[74],"LLM-NPU,":[75],"a":[76],"comprehensive":[77],"software-hardware":[78],"co-optimization":[79],"framework":[80],"that":[81,118],"enables":[82],"scalable,":[83],"power-efficient":[84],"LLM":[85,126,173],"deployment":[86],"on":[87],"under":[89],"tight":[90],"compute":[91,123],"memory":[93,120],"budgets.":[94],"We":[95],"present":[96],"solutions":[98],"vertical":[101],"horizontal":[103],"operator":[104],"fusion,":[105],"quantization-aware":[106],"weight":[107],"compression,":[108],"hybrid":[109],"key-value":[110],"(KV)":[111],"quantization,":[112],"eviction":[113],"strategies,":[114],"static-shape":[116],"inference":[117],"target":[119],"bottlenecks":[121],"inefficiencies":[124],"execution.":[127],"On":[128],"hardware":[130],"side,":[131],"we":[132],"explore":[133],"domain-specialized":[134],"NPU":[135],"enhancements,":[136],"including":[137],"processing-in-memory":[138],"architectures,":[139],"extended":[140],"input":[141],"channel":[142],"accumulation,":[143],"structured":[144],"sparsity":[145],"acceleration,":[146],"GEMM":[147],"engine":[148],"optimizations,":[149],"mixed":[150],"precision,":[151],"microscaling":[152],"format":[153],"support,":[154],"fusion-aware":[156],"execution":[157],"pipelines.":[158],"These":[159],"co-designed":[160],"innovations":[161],"can":[162],"collectively":[163],"energy,":[166],"throughput,":[167],"latency":[169],"workloads.":[174]},"counts_by_year":[{"year":2025,"cited_by_count":2}],"updated_date":"2026-07-29T14:22:42.915294","created_date":"2025-10-10T00:00:00"}