{"id":"https://openalex.org/W7162426210","doi":"https://doi.org/10.1109/ispass69572.2026.00053","title":"HeadSkip: Characterizing and Accelerating \u2018Small\u2019 Language Models on Edge CPUs","display_name":"HeadSkip: Characterizing and Accelerating \u2018Small\u2019 Language Models on Edge CPUs","publication_year":2026,"publication_date":"2026-04-26","ids":{"openalex":"https://openalex.org/W7162426210","doi":"https://doi.org/10.1109/ispass69572.2026.00053"},"language":null,"primary_location":{"id":"doi:10.1109/ispass69572.2026.00053","is_oa":false,"landing_page_url":"https://doi.org/10.1109/ispass69572.2026.00053","pdf_url":null,"source":null,"license":null,"license_id":null,"version":"publishedVersion","is_accepted":true,"is_published":true,"raw_source_name":"2026 IEEE International Symposium on Performance Analysis of Systems and Software (ISPASS)","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/A5136517176","display_name":"Ian Sartor","orcid":null},"institutions":[{"id":"https://openalex.org/I75717288","display_name":"Rogers (United States)","ror":"https://ror.org/05m9vrv91","country_code":"US","type":"company","lineage":["https://openalex.org/I75717288"]}],"countries":["US"],"is_corresponding":false,"raw_author_name":"Ian Sartor","raw_affiliation_strings":["University of Toronto,The Edward S. Rogers Sr. Department of Electrical and Computer Engineering,Toronto,ON"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"University of Toronto,The Edward S. Rogers Sr. Department of Electrical and Computer Engineering,Toronto,ON","institution_ids":["https://openalex.org/I75717288"]}]},{"author_position":"middle","author":{"id":"https://openalex.org/A5137084653","display_name":"Karthik Ganesan","orcid":null},"institutions":[{"id":"https://openalex.org/I75717288","display_name":"Rogers (United States)","ror":"https://ror.org/05m9vrv91","country_code":"US","type":"company","lineage":["https://openalex.org/I75717288"]}],"countries":["US"],"is_corresponding":false,"raw_author_name":"Karthik Ganesan","raw_affiliation_strings":["University of Toronto,The Edward S. Rogers Sr. Department of Electrical and Computer Engineering,Toronto,ON"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"University of Toronto,The Edward S. Rogers Sr. Department of Electrical and Computer Engineering,Toronto,ON","institution_ids":["https://openalex.org/I75717288"]}]},{"author_position":"middle","author":{"id":"https://openalex.org/A5132939603","display_name":"Anh Nguyen","orcid":null},"institutions":[{"id":"https://openalex.org/I2252096349","display_name":"Fujitsu (Japan)","ror":"https://ror.org/038e2g226","country_code":"JP","type":"company","lineage":["https://openalex.org/I2252096349"]},{"id":"https://openalex.org/I4210159607","display_name":"Fujitsu (China)","ror":"https://ror.org/04w4yzw62","country_code":"CN","type":"company","lineage":["https://openalex.org/I4210159607"]}],"countries":["CN","JP"],"is_corresponding":false,"raw_author_name":"Anh Nguyen","raw_affiliation_strings":["Fujitsu Limited,Toronto,ON"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"Fujitsu Limited,Toronto,ON","institution_ids":["https://openalex.org/I2252096349","https://openalex.org/I4210159607"]}]},{"author_position":"last","author":{"id":"https://openalex.org/A5072544251","display_name":"Andreas Moshovos","orcid":"https://orcid.org/0000-0001-7768-367X"},"institutions":[{"id":"https://openalex.org/I75717288","display_name":"Rogers (United States)","ror":"https://ror.org/05m9vrv91","country_code":"US","type":"company","lineage":["https://openalex.org/I75717288"]}],"countries":["US"],"is_corresponding":false,"raw_author_name":"Andreas Moshovos","raw_affiliation_strings":["University of Toronto,The Edward S. Rogers Sr. Department of Electrical and Computer Engineering,Toronto,ON"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"University of Toronto,The Edward S. Rogers Sr. Department of Electrical and Computer Engineering,Toronto,ON","institution_ids":["https://openalex.org/I75717288"]}]}],"institutions":[],"countries_distinct_count":3,"institutions_distinct_count":3,"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":"484","last_page":"494"},"is_retracted":false,"is_paratext":false,"is_xpac":false,"primary_topic":{"id":"https://openalex.org/T10054","display_name":"Parallel Computing and Optimization Techniques","score":0.24420000612735748,"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":0.24420000612735748,"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/T10181","display_name":"Natural Language Processing Techniques","score":0.2142000049352646,"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/T11450","display_name":"Model-Driven Software Engineering Techniques","score":0.025599999353289604,"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"}}],"keywords":[{"id":"https://openalex.org/keywords/enhanced-data-rates-for-gsm-evolution","display_name":"Enhanced Data Rates for GSM Evolution","score":0.4332999885082245},{"id":"https://openalex.org/keywords/abstraction","display_name":"Abstraction","score":0.29440000653266907},{"id":"https://openalex.org/keywords/modeling-language","display_name":"Modeling language","score":0.26899999380111694},{"id":"https://openalex.org/keywords/natural-language","display_name":"Natural language","score":0.2671999931335449},{"id":"https://openalex.org/keywords/coupling","display_name":"Coupling (piping)","score":0.26159998774528503}],"concepts":[{"id":"https://openalex.org/C41008148","wikidata":"https://www.wikidata.org/wiki/Q21198","display_name":"Computer science","level":0,"score":0.6366000175476074},{"id":"https://openalex.org/C162307627","wikidata":"https://www.wikidata.org/wiki/Q204833","display_name":"Enhanced Data Rates for GSM Evolution","level":2,"score":0.4332999885082245},{"id":"https://openalex.org/C199360897","wikidata":"https://www.wikidata.org/wiki/Q9143","display_name":"Programming language","level":1,"score":0.4081000089645386},{"id":"https://openalex.org/C11413529","wikidata":"https://www.wikidata.org/wiki/Q8366","display_name":"Algorithm","level":1,"score":0.3499000072479248},{"id":"https://openalex.org/C154945302","wikidata":"https://www.wikidata.org/wiki/Q11660","display_name":"Artificial intelligence","level":1,"score":0.2976999878883362},{"id":"https://openalex.org/C124304363","wikidata":"https://www.wikidata.org/wiki/Q673661","display_name":"Abstraction","level":2,"score":0.29440000653266907},{"id":"https://openalex.org/C80444323","wikidata":"https://www.wikidata.org/wiki/Q2878974","display_name":"Theoretical computer science","level":1,"score":0.29330000281333923},{"id":"https://openalex.org/C121684516","wikidata":"https://www.wikidata.org/wiki/Q7600677","display_name":"Computer graphics (images)","level":1,"score":0.27410000562667847},{"id":"https://openalex.org/C179603123","wikidata":"https://www.wikidata.org/wiki/Q1941921","display_name":"Modeling language","level":3,"score":0.26899999380111694},{"id":"https://openalex.org/C195324797","wikidata":"https://www.wikidata.org/wiki/Q33742","display_name":"Natural language","level":2,"score":0.2671999931335449},{"id":"https://openalex.org/C131584629","wikidata":"https://www.wikidata.org/wiki/Q4308705","display_name":"Coupling (piping)","level":2,"score":0.26159998774528503},{"id":"https://openalex.org/C137293760","wikidata":"https://www.wikidata.org/wiki/Q3621696","display_name":"Language model","level":2,"score":0.2526000142097473}],"mesh":[],"locations_count":1,"locations":[{"id":"doi:10.1109/ispass69572.2026.00053","is_oa":false,"landing_page_url":"https://doi.org/10.1109/ispass69572.2026.00053","pdf_url":null,"source":null,"license":null,"license_id":null,"version":"publishedVersion","is_accepted":true,"is_published":true,"raw_source_name":"2026 IEEE International Symposium on Performance Analysis of Systems and Software (ISPASS)","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":11,"referenced_works":["https://openalex.org/W2048176942","https://openalex.org/W2962677625","https://openalex.org/W4221164185","https://openalex.org/W4385245566","https://openalex.org/W4389523854","https://openalex.org/W4401042553","https://openalex.org/W4406207083","https://openalex.org/W4414638244","https://openalex.org/W4415799217","https://openalex.org/W7133219248","https://openalex.org/W7138830883"],"related_works":[],"abstract_inverted_index":{"Recent":[0],"years":[1],"have":[2,31],"seen":[3,149],"a":[4,88,162,168,188,224],"surge":[5],"in":[6,18,91,139,141,161,202,218,253],"the":[7,61,71,80,83,92,101,130,136,156,173,176,182,185,198,232,254],"popularity":[8],"of":[9,70,82,184,200,231],"large":[10,108],"language":[11,34],"models.":[12],"Such":[13],"models":[14,35],"are":[15,211],"typically":[16],"run":[17,39],"datacenters":[19],"using":[20],"high-end":[21],"GPUs.":[22],"However,":[23,204],"due":[24],"to":[25,38,67,87,128,146,265],"latency":[26],"and":[27,52,58,164,179],"privacy":[28],"concerns,":[29],"companies":[30],"proposed":[32],"\u2018small\u2019":[33],"(SLMs),":[36],"designed":[37],"locally":[40],"on":[41,54,119,197,268,273],"edge":[42,56,275],"devices.":[43],"We":[44,191],"characterize":[45],"two":[46,125,274],"notable":[47],"SLMs":[48,111],"(i.e.,":[49],"Gemma":[50,116],"3":[51],"Qwen3)":[53],"an":[55],"CPU":[57],"find":[59],"that":[60],"final":[62,77,131,137],"layer":[63,78,138,233],"accounts":[64],"for":[65,110,115,159],"up":[66],"$31.5":[68],"\\%$":[69],"total":[72],"SLM":[73],"inference":[74],"runtime.":[75],"This":[76,257],"maps":[79],"output":[81,158,172],"last":[84],"decoder":[85],"block":[86],"specific":[89],"token":[90,174],"model\u2019s":[93],"vocabulary.":[94],"Therefore,":[95],"this":[96,120,193,205,237],"layer\u2019s":[97],"runtime":[98],"scales":[99],"with":[100,175],"vocabulary":[102],"size,":[103],"which":[104,210,251],"can":[105],"be":[106],"very":[107],"even":[109,261],"(e.g.,":[112],"262k":[113],"tokens":[114,160,250],"3).":[117],"Building":[118],"observation,":[121],"we":[122,134,152,154,171,235],"present":[123],"HeadSkip,":[124],"complimentary":[126],"techniques":[127],"speed-up":[129],"layer.":[132],"First,":[133],"execute":[135,229],"chunks":[140],"order":[142,226],"from":[143],"most":[144,230],"frequently":[145,148,217],"least":[147],"tokens.":[150],"As":[151],"run,":[153],"check":[155,236],"logit":[157,178],"chunk":[163],"if":[165],"any":[166],"exceed":[167],"pre-set":[169],"threshold,":[170],"largest":[177],"skip":[180],"executing":[181],"rest":[183],"layer,":[186],"obtaining":[187],"significant":[189],"speedup.":[190],"perform":[192],"re-ordering":[194],"statically":[195],"based":[196],"frequency":[199],"words":[201,209],"English.":[203],"does":[206],"not":[207],"capture":[208],"generally":[212],"rare":[213],"but":[214],"may":[215,227],"appear":[216,252],"some":[219],"contexts.":[220],"For":[221],"such":[222],"words,":[223],"static":[225],"still":[228],"until":[234],"\u2018infrequent\u2019":[238],"token,":[239],"diminishing":[240],"our":[241],"speedup":[242],"gains.":[243],"To":[244],"avoid":[245],"this,":[246],"HeadSkip":[247],"also":[248],"caches":[249],"current":[255],"context.":[256],"helps":[258],"us":[259],"achieve":[260],"greater":[262],"speed-ups":[263],"(up":[264],"$45":[266],"\\%$)":[267],"modern":[269],"SLMs,":[270],"when":[271],"evaluated":[272],"CPUs.":[276]},"counts_by_year":[],"updated_date":"2026-07-22T07:51:19.307946","created_date":"2026-05-27T00:00:00"}