{"id":"https://openalex.org/W7155550404","doi":"https://doi.org/10.1145/3767295.3769382","title":"Scaling LLM Test-Time Compute with Mobile NPU on Smartphones","display_name":"Scaling LLM Test-Time Compute with Mobile NPU on Smartphones","publication_year":2026,"publication_date":"2026-04-24","ids":{"openalex":"https://openalex.org/W7155550404","doi":"https://doi.org/10.1145/3767295.3769382"},"language":null,"primary_location":{"id":"doi:10.1145/3767295.3769382","is_oa":true,"landing_page_url":"https://doi.org/10.1145/3767295.3769382","pdf_url":null,"source":null,"license":"cc-by","license_id":"https://openalex.org/licenses/cc-by","version":"publishedVersion","is_accepted":true,"is_published":true,"raw_source_name":"Proceedings of the 21st European Conference on Computer Systems","raw_type":"proceedings-article"},"type":"article","indexed_in":["crossref"],"open_access":{"is_oa":true,"oa_status":"gold","oa_url":"https://doi.org/10.1145/3767295.3769382","any_repository_has_fulltext":null},"authorships":[{"author_position":"first","author":{"id":"https://openalex.org/A5093259788","display_name":"Zixu Hao","orcid":"https://orcid.org/0009-0007-1671-1367"},"institutions":[{"id":"https://openalex.org/I99065089","display_name":"Tsinghua University","ror":"https://ror.org/03cve4549","country_code":"CN","type":"education","lineage":["https://openalex.org/I99065089"]}],"countries":["CN"],"is_corresponding":false,"raw_author_name":"Zixu Hao","raw_affiliation_strings":["Tsinghua University, Beijing, China"],"raw_orcid":"https://orcid.org/0009-0007-1671-1367","affiliations":[{"raw_affiliation_string":"Tsinghua University, Beijing, China","institution_ids":["https://openalex.org/I99065089"]}]},{"author_position":"middle","author":{"id":"https://openalex.org/A5071351577","display_name":"Jianyu Wei","orcid":"https://orcid.org/0009-0009-0830-044X"},"institutions":[{"id":"https://openalex.org/I126520041","display_name":"University of Science and Technology of China","ror":"https://ror.org/04c4dkn09","country_code":"CN","type":"education","lineage":["https://openalex.org/I126520041","https://openalex.org/I19820366"]}],"countries":["CN"],"is_corresponding":false,"raw_author_name":"Jianyu Wei","raw_affiliation_strings":["University of Science and Technology of China, Hefei, China"],"raw_orcid":"https://orcid.org/0009-0009-0830-044X","affiliations":[{"raw_affiliation_string":"University of Science and Technology of China, Hefei, China","institution_ids":["https://openalex.org/I126520041"]}]},{"author_position":"middle","author":{"id":"https://openalex.org/A5067496010","display_name":"Tuowei Wang","orcid":"https://orcid.org/0009-0006-8272-8151"},"institutions":[{"id":"https://openalex.org/I99065089","display_name":"Tsinghua University","ror":"https://ror.org/03cve4549","country_code":"CN","type":"education","lineage":["https://openalex.org/I99065089"]}],"countries":["CN"],"is_corresponding":false,"raw_author_name":"Tuowei Wang","raw_affiliation_strings":["Tsinghua University, Beijing, China"],"raw_orcid":"https://orcid.org/0009-0006-8272-8151","affiliations":[{"raw_affiliation_string":"Tsinghua University, Beijing, China","institution_ids":["https://openalex.org/I99065089"]}]},{"author_position":"middle","author":{"id":"https://openalex.org/A5134470748","display_name":"Minxing Huang","orcid":"https://orcid.org/0009-0003-8024-4928"},"institutions":[{"id":"https://openalex.org/I99065089","display_name":"Tsinghua University","ror":"https://ror.org/03cve4549","country_code":"CN","type":"education","lineage":["https://openalex.org/I99065089"]}],"countries":["CN"],"is_corresponding":false,"raw_author_name":"Minxing Huang","raw_affiliation_strings":["Tsinghua University, Beijing, China"],"raw_orcid":"https://orcid.org/0009-0003-8024-4928","affiliations":[{"raw_affiliation_string":"Tsinghua University, Beijing, China","institution_ids":["https://openalex.org/I99065089"]}]},{"author_position":"middle","author":{"id":"https://openalex.org/A5070156365","display_name":"Huiqiang Jiang","orcid":"https://orcid.org/0000-0002-1327-4882"},"institutions":[{"id":"https://openalex.org/I4210113369","display_name":"Microsoft Research Asia (China)","ror":"https://ror.org/0300m5276","country_code":"CN","type":"company","lineage":["https://openalex.org/I1290206253","https://openalex.org/I4210113369"]}],"countries":["CN"],"is_corresponding":false,"raw_author_name":"Huiqiang Jiang","raw_affiliation_strings":["Microsoft Research, Shanghai, China"],"raw_orcid":"https://orcid.org/0000-0002-1327-4882","affiliations":[{"raw_affiliation_string":"Microsoft Research, Shanghai, China","institution_ids":["https://openalex.org/I4210113369"]}]},{"author_position":"middle","author":{"id":"https://openalex.org/A5101989690","display_name":"Shiqi Jiang","orcid":"https://orcid.org/0000-0002-4685-9633"},"institutions":[{"id":"https://openalex.org/I4210113369","display_name":"Microsoft Research Asia (China)","ror":"https://ror.org/0300m5276","country_code":"CN","type":"company","lineage":["https://openalex.org/I1290206253","https://openalex.org/I4210113369"]}],"countries":["CN"],"is_corresponding":false,"raw_author_name":"Shiqi Jiang","raw_affiliation_strings":["Microsoft Research, Shanghai, China"],"raw_orcid":"https://orcid.org/0000-0002-4685-9633","affiliations":[{"raw_affiliation_string":"Microsoft Research, Shanghai, China","institution_ids":["https://openalex.org/I4210113369"]}]},{"author_position":"middle","author":{"id":"https://openalex.org/A5101534161","display_name":"Ting Cao","orcid":"https://orcid.org/0000-0002-9107-013X"},"institutions":[{"id":"https://openalex.org/I99065089","display_name":"Tsinghua University","ror":"https://ror.org/03cve4549","country_code":"CN","type":"education","lineage":["https://openalex.org/I99065089"]}],"countries":["CN"],"is_corresponding":false,"raw_author_name":"Ting Cao","raw_affiliation_strings":["Institute for AI Industry Research (AIR), Tsinghua University, Beijing, China"],"raw_orcid":"https://orcid.org/0000-0002-9107-013X","affiliations":[{"raw_affiliation_string":"Institute for AI Industry Research (AIR), Tsinghua University, Beijing, China","institution_ids":["https://openalex.org/I99065089"]}]},{"author_position":"last","author":{"id":"https://openalex.org/A5004848767","display_name":"Ju Ren","orcid":null},"institutions":[{"id":"https://openalex.org/I99065089","display_name":"Tsinghua University","ror":"https://ror.org/03cve4549","country_code":"CN","type":"education","lineage":["https://openalex.org/I99065089"]}],"countries":["CN"],"is_corresponding":false,"raw_author_name":"Ju Ren","raw_affiliation_strings":["Tsinghua University, Beijing, China"],"raw_orcid":"https://orcid.org/0000-0003-2782-183X","affiliations":[{"raw_affiliation_string":"Tsinghua University, Beijing, China","institution_ids":["https://openalex.org/I99065089"]}]}],"institutions":[],"countries_distinct_count":1,"institutions_distinct_count":8,"corresponding_author_ids":[],"corresponding_institution_ids":[],"apc_list":null,"apc_paid":null,"fwci":0.0,"has_fulltext":false,"cited_by_count":0,"citation_normalized_percentile":{"value":0.65627591,"is_in_top_1_percent":false,"is_in_top_10_percent":false},"cited_by_percentile_year":null,"biblio":{"volume":null,"issue":null,"first_page":"2157","last_page":"2172"},"is_retracted":false,"is_paratext":false,"is_xpac":false,"primary_topic":{"id":"https://openalex.org/T10904","display_name":"Embedded Systems Design Techniques","score":0.06960000097751617,"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/T10904","display_name":"Embedded Systems Design Techniques","score":0.06960000097751617,"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/T10743","display_name":"Software Testing and Debugging Techniques","score":0.06650000065565109,"subfield":{"id":"https://openalex.org/subfields/1712","display_name":"Software"},"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/T12127","display_name":"Software System Performance and Reliability","score":0.060600001364946365,"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/scaling","display_name":"Scaling","score":0.45910000801086426},{"id":"https://openalex.org/keywords/mobile-device","display_name":"Mobile device","score":0.35850000381469727},{"id":"https://openalex.org/keywords/field","display_name":"Field (mathematics)","score":0.32519999146461487},{"id":"https://openalex.org/keywords/scale","display_name":"Scale (ratio)","score":0.2847999930381775},{"id":"https://openalex.org/keywords/mobile-computing","display_name":"Mobile computing","score":0.2800000011920929}],"concepts":[{"id":"https://openalex.org/C41008148","wikidata":"https://www.wikidata.org/wiki/Q21198","display_name":"Computer science","level":0,"score":0.5809999704360962},{"id":"https://openalex.org/C99844830","wikidata":"https://www.wikidata.org/wiki/Q102441924","display_name":"Scaling","level":2,"score":0.45910000801086426},{"id":"https://openalex.org/C186967261","wikidata":"https://www.wikidata.org/wiki/Q5082128","display_name":"Mobile device","level":2,"score":0.35850000381469727},{"id":"https://openalex.org/C9652623","wikidata":"https://www.wikidata.org/wiki/Q190109","display_name":"Field (mathematics)","level":2,"score":0.32519999146461487},{"id":"https://openalex.org/C79403827","wikidata":"https://www.wikidata.org/wiki/Q3988","display_name":"Real-time computing","level":1,"score":0.30880001187324524},{"id":"https://openalex.org/C2778755073","wikidata":"https://www.wikidata.org/wiki/Q10858537","display_name":"Scale (ratio)","level":2,"score":0.2847999930381775},{"id":"https://openalex.org/C144543869","wikidata":"https://www.wikidata.org/wiki/Q2738570","display_name":"Mobile computing","level":2,"score":0.2800000011920929},{"id":"https://openalex.org/C204323151","wikidata":"https://www.wikidata.org/wiki/Q905424","display_name":"Range (aeronautics)","level":2,"score":0.2612000107765198},{"id":"https://openalex.org/C99498987","wikidata":"https://www.wikidata.org/wiki/Q2210247","display_name":"Noise (video)","level":3,"score":0.26010000705718994},{"id":"https://openalex.org/C44154836","wikidata":"https://www.wikidata.org/wiki/Q45045","display_name":"Simulation","level":1,"score":0.24940000474452972}],"mesh":[],"locations_count":1,"locations":[{"id":"doi:10.1145/3767295.3769382","is_oa":true,"landing_page_url":"https://doi.org/10.1145/3767295.3769382","pdf_url":null,"source":null,"license":"cc-by","license_id":"https://openalex.org/licenses/cc-by","version":"publishedVersion","is_accepted":true,"is_published":true,"raw_source_name":"Proceedings of the 21st European Conference on Computer Systems","raw_type":"proceedings-article"}],"best_oa_location":{"id":"doi:10.1145/3767295.3769382","is_oa":true,"landing_page_url":"https://doi.org/10.1145/3767295.3769382","pdf_url":null,"source":null,"license":"cc-by","license_id":"https://openalex.org/licenses/cc-by","version":"publishedVersion","is_accepted":true,"is_published":true,"raw_source_name":"Proceedings of the 21st European Conference on Computer Systems","raw_type":"proceedings-article"},"sustainable_development_goals":[{"id":"https://metadata.un.org/sdg/7","display_name":"Affordable and clean energy","score":0.5946401953697205}],"awards":[],"funders":[],"has_content":{"pdf":false,"grobid_xml":false},"content_urls":null,"referenced_works_count":0,"referenced_works":[],"related_works":[],"abstract_inverted_index":{"Deploying":[0],"Large":[1],"Language":[2],"Models":[3],"(LLMs)":[4],"on":[5,59,141],"mobile":[6,28,60],"devices":[7],"faces":[8],"the":[9,64,133,176],"challenge":[10],"of":[11,66,178],"insufficient":[12],"performance":[13,65],"in":[14,21,86],"smaller":[15,67,167],"models":[16,168],"and":[17,83,111,121,125,158],"excessive":[18],"resource":[19],"consumption":[20],"larger":[22,179],"ones.":[23],"This":[24],"paper":[25],"highlights":[26],"that":[27,102,131,166],"Neural":[29],"Processing":[30],"Units":[31],"(NPUs)":[32],"have":[33],"underutilized":[34],"computational":[35],"resources,":[36],"particularly":[37],"their":[38],"matrix":[39],"multiplication":[40],"units,":[41],"during":[42],"typical":[43],"LLM":[44],"inference.":[45],"To":[46,89],"leverage":[47],"this":[48,70],"wasted":[49],"compute":[50,135],"capacity,":[51],"we":[52,92,164],"propose":[53],"applying":[54],"parallel":[55],"test-time":[56,139,170],"scaling":[57,140,171],"techniques":[58],"NPUs":[61],"to":[62,137,153],"enhance":[63],"LLMs.":[68],"However,":[69],"approach":[71,148],"confronts":[72],"inherent":[73],"NPU":[74,107],"challenges,":[75],"including":[76],"inadequate":[77],"hardware":[78],"support":[79,138],"for":[80,115,155,160],"fine-grained":[81],"quantization":[82,100,105],"low":[84],"efficiency":[85],"general-purpose":[87],"computations.":[88],"overcome":[90],"these,":[91],"introduce":[93],"two":[94],"key":[95],"techniques:":[96],"a":[97,182],"hardware-aware":[98],"tile":[99],"scheme":[101],"aligns":[103],"group":[104],"with":[106],"memory":[108],"access":[109],"patterns,":[110],"efficient":[112],"LUT-based":[113],"replacements":[114],"complex":[116],"operations":[117],"such":[118],"as":[119],"Softmax":[120],"dequantization.":[122],"We":[123],"design":[124],"implement":[126],"an":[127],"end-to-end":[128],"inference":[129],"system":[130],"leverages":[132],"NPU's":[134],"capability":[136],"Qualcomm":[142],"Snapdragon":[143],"platforms.":[144],"Experiments":[145],"show":[146],"our":[147],"brings":[149],"significant":[150],"speedups:":[151],"up":[152],"19.0\u00d7":[154],"mixed-precision":[156],"GEMM":[157],"2.2\u00d7":[159],"Softmax.":[161],"More":[162],"importantly,":[163],"demonstrate":[165],"using":[169],"can":[172],"match":[173],"or":[174],"exceed":[175],"accuracy":[177],"models,":[180],"achieving":[181],"new":[183],"performance-cost":[184],"Pareto":[185],"frontier.":[186]},"counts_by_year":[],"updated_date":"2026-06-11T09:08:48.828518","created_date":"2026-04-25T00:00:00"}