{"id":"https://openalex.org/W4415593734","doi":"https://doi.org/10.1109/access.2025.3625771","title":"Sub 4-bit Power-of-Two-Based Mixed-Precision Quantization for Efficient LLM Compression and Acceleration","display_name":"Sub 4-bit Power-of-Two-Based Mixed-Precision Quantization for Efficient LLM Compression and Acceleration","publication_year":2025,"publication_date":"2025-01-01","ids":{"openalex":"https://openalex.org/W4415593734","doi":"https://doi.org/10.1109/access.2025.3625771"},"language":null,"primary_location":{"id":"doi:10.1109/access.2025.3625771","is_oa":true,"landing_page_url":"https://doi.org/10.1109/access.2025.3625771","pdf_url":null,"source":{"id":"https://openalex.org/S2485537415","display_name":"IEEE Access","issn_l":"2169-3536","issn":["2169-3536"],"is_oa":true,"is_in_doaj":true,"is_core":true,"host_organization":"https://openalex.org/P4310319808","host_organization_name":"Institute of Electrical and Electronics Engineers","host_organization_lineage":["https://openalex.org/P4310319808"],"host_organization_lineage_names":["Institute of Electrical and Electronics Engineers"],"type":"journal"},"license":"cc-by","license_id":"https://openalex.org/licenses/cc-by","version":"publishedVersion","is_accepted":true,"is_published":true,"raw_source_name":"IEEE Access","raw_type":"journal-article"},"type":"article","indexed_in":["crossref","doaj"],"open_access":{"is_oa":true,"oa_status":"gold","oa_url":"https://doi.org/10.1109/access.2025.3625771","any_repository_has_fulltext":false},"authorships":[{"author_position":"first","author":{"id":"https://openalex.org/A5019082108","display_name":"Han Cho","orcid":"https://orcid.org/0009-0000-9010-7888"},"institutions":[{"id":"https://openalex.org/I130701444","display_name":"Georgia Institute of Technology","ror":"https://ror.org/01zkghx44","country_code":"US","type":"education","lineage":["https://openalex.org/I130701444"]}],"countries":["US"],"is_corresponding":false,"raw_author_name":"Han Cho","raw_affiliation_strings":["School of Electrical and Computer Engineering, Georgia Institute of Technology, Atlanta, GA, USA"],"raw_orcid":"https://orcid.org/0009-0000-9010-7888","affiliations":[{"raw_affiliation_string":"School of Electrical and Computer Engineering, Georgia Institute of Technology, Atlanta, GA, USA","institution_ids":["https://openalex.org/I130701444"]}]},{"author_position":"middle","author":{"id":"https://openalex.org/A5120152261","display_name":"Apurba Prasad Padhy","orcid":"https://orcid.org/0009-0008-0401-3809"},"institutions":[{"id":"https://openalex.org/I130701444","display_name":"Georgia Institute of Technology","ror":"https://ror.org/01zkghx44","country_code":"US","type":"education","lineage":["https://openalex.org/I130701444"]}],"countries":["US"],"is_corresponding":false,"raw_author_name":"Apurba Prasad Padhy","raw_affiliation_strings":["School of Electrical and Computer Engineering, Georgia Institute of Technology, Atlanta, GA, USA"],"raw_orcid":"https://orcid.org/0009-0008-0401-3809","affiliations":[{"raw_affiliation_string":"School of Electrical and Computer Engineering, Georgia Institute of Technology, Atlanta, GA, USA","institution_ids":["https://openalex.org/I130701444"]}]},{"author_position":"middle","author":{"id":"https://openalex.org/A5119152038","display_name":"Fernando Camacho","orcid":null},"institutions":[{"id":"https://openalex.org/I4210113003","display_name":"Physical Sciences (United States)","ror":"https://ror.org/021qvjc46","country_code":"US","type":"company","lineage":["https://openalex.org/I4210113003"]}],"countries":["US"],"is_corresponding":false,"raw_author_name":"Fernando Camacho","raw_affiliation_strings":["Laboratory for Physical Sciences, College Park, MD, USA"],"raw_orcid":"https://orcid.org/0009-0006-3181-4076","affiliations":[{"raw_affiliation_string":"Laboratory for Physical Sciences, College Park, MD, USA","institution_ids":["https://openalex.org/I4210113003"]}]},{"author_position":"last","author":{"id":"https://openalex.org/A5009591041","display_name":"Saibal Mukhopadhyay","orcid":"https://orcid.org/0000-0002-8894-3390"},"institutions":[{"id":"https://openalex.org/I130701444","display_name":"Georgia Institute of Technology","ror":"https://ror.org/01zkghx44","country_code":"US","type":"education","lineage":["https://openalex.org/I130701444"]}],"countries":["US"],"is_corresponding":false,"raw_author_name":"Saibal Mukhopadhyay","raw_affiliation_strings":["School of Electrical and Computer Engineering, Georgia Institute of Technology, Atlanta, GA, USA"],"raw_orcid":"https://orcid.org/0000-0002-8894-3390","affiliations":[{"raw_affiliation_string":"School of Electrical and Computer Engineering, Georgia Institute of Technology, Atlanta, GA, USA","institution_ids":["https://openalex.org/I130701444"]}]}],"institutions":[],"countries_distinct_count":1,"institutions_distinct_count":2,"corresponding_author_ids":[],"corresponding_institution_ids":[],"apc_list":{"value":1850,"currency":"USD","value_usd":1850},"apc_paid":{"value":1850,"currency":"USD","value_usd":1850},"fwci":0.0,"has_fulltext":false,"cited_by_count":0,"citation_normalized_percentile":{"value":0.2726979,"is_in_top_1_percent":false,"is_in_top_10_percent":false},"cited_by_percentile_year":null,"biblio":{"volume":"13","issue":null,"first_page":"209356","last_page":"209367"},"is_retracted":false,"is_paratext":false,"is_xpac":false,"primary_topic":{"id":"https://openalex.org/T10901","display_name":"Advanced Data Compression Techniques","score":0.9883999824523926,"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/T10901","display_name":"Advanced Data Compression Techniques","score":0.9883999824523926,"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/T10125","display_name":"Advanced Wireless Communication Techniques","score":0.9768999814987183,"subfield":{"id":"https://openalex.org/subfields/2208","display_name":"Electrical and Electronic Engineering"},"field":{"id":"https://openalex.org/fields/22","display_name":"Engineering"},"domain":{"id":"https://openalex.org/domains/3","display_name":"Physical Sciences"}},{"id":"https://openalex.org/T11034","display_name":"Digital Filter Design and Implementation","score":0.9710999727249146,"subfield":{"id":"https://openalex.org/subfields/1711","display_name":"Signal Processing"},"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/datapath","display_name":"Datapath","score":0.8489000201225281},{"id":"https://openalex.org/keywords/quantization","display_name":"Quantization (signal processing)","score":0.8003000020980835},{"id":"https://openalex.org/keywords/hardware-acceleration","display_name":"Hardware acceleration","score":0.49230000376701355},{"id":"https://openalex.org/keywords/logarithm","display_name":"Logarithm","score":0.4587000012397766},{"id":"https://openalex.org/keywords/matrix-multiplication","display_name":"Matrix multiplication","score":0.42809998989105225},{"id":"https://openalex.org/keywords/software-deployment","display_name":"Software deployment","score":0.41110000014305115},{"id":"https://openalex.org/keywords/edge-device","display_name":"Edge device","score":0.4016000032424927},{"id":"https://openalex.org/keywords/software","display_name":"Software","score":0.3926999866962433}],"concepts":[{"id":"https://openalex.org/C2781198647","wikidata":"https://www.wikidata.org/wiki/Q1633673","display_name":"Datapath","level":2,"score":0.8489000201225281},{"id":"https://openalex.org/C28855332","wikidata":"https://www.wikidata.org/wiki/Q198099","display_name":"Quantization (signal processing)","level":2,"score":0.8003000020980835},{"id":"https://openalex.org/C41008148","wikidata":"https://www.wikidata.org/wiki/Q21198","display_name":"Computer science","level":0,"score":0.7613000273704529},{"id":"https://openalex.org/C13164978","wikidata":"https://www.wikidata.org/wiki/Q600158","display_name":"Hardware acceleration","level":3,"score":0.49230000376701355},{"id":"https://openalex.org/C39927690","wikidata":"https://www.wikidata.org/wiki/Q11197","display_name":"Logarithm","level":2,"score":0.4587000012397766},{"id":"https://openalex.org/C17349429","wikidata":"https://www.wikidata.org/wiki/Q1049914","display_name":"Matrix multiplication","level":3,"score":0.42809998989105225},{"id":"https://openalex.org/C105339364","wikidata":"https://www.wikidata.org/wiki/Q2297740","display_name":"Software deployment","level":2,"score":0.41110000014305115},{"id":"https://openalex.org/C138236772","wikidata":"https://www.wikidata.org/wiki/Q25098575","display_name":"Edge device","level":3,"score":0.4016000032424927},{"id":"https://openalex.org/C2777904410","wikidata":"https://www.wikidata.org/wiki/Q7397","display_name":"Software","level":2,"score":0.3926999866962433},{"id":"https://openalex.org/C2780595030","wikidata":"https://www.wikidata.org/wiki/Q3860309","display_name":"Multiplication (music)","level":2,"score":0.39160001277923584},{"id":"https://openalex.org/C9390403","wikidata":"https://www.wikidata.org/wiki/Q3966","display_name":"Computer hardware","level":1,"score":0.38989999890327454},{"id":"https://openalex.org/C173608175","wikidata":"https://www.wikidata.org/wiki/Q232661","display_name":"Parallel computing","level":1,"score":0.3887999951839447},{"id":"https://openalex.org/C11413529","wikidata":"https://www.wikidata.org/wiki/Q8366","display_name":"Algorithm","level":1,"score":0.3865000009536743},{"id":"https://openalex.org/C111335779","wikidata":"https://www.wikidata.org/wiki/Q3454686","display_name":"Reduction (mathematics)","level":2,"score":0.3528999984264374},{"id":"https://openalex.org/C113775141","wikidata":"https://www.wikidata.org/wiki/Q428691","display_name":"Computer engineering","level":1,"score":0.34470000863075256},{"id":"https://openalex.org/C2742236","wikidata":"https://www.wikidata.org/wiki/Q924713","display_name":"Efficient energy use","level":2,"score":0.310699999332428},{"id":"https://openalex.org/C106487976","wikidata":"https://www.wikidata.org/wiki/Q685816","display_name":"Matrix (chemical analysis)","level":2,"score":0.30160000920295715},{"id":"https://openalex.org/C162307627","wikidata":"https://www.wikidata.org/wiki/Q204833","display_name":"Enhanced Data Rates for GSM Evolution","level":2,"score":0.287200003862381},{"id":"https://openalex.org/C149635348","wikidata":"https://www.wikidata.org/wiki/Q193040","display_name":"Embedded system","level":1,"score":0.2806999981403351},{"id":"https://openalex.org/C459310","wikidata":"https://www.wikidata.org/wiki/Q117801","display_name":"Computational science","level":1,"score":0.2768000066280365},{"id":"https://openalex.org/C117896860","wikidata":"https://www.wikidata.org/wiki/Q11376","display_name":"Acceleration","level":2,"score":0.27070000767707825},{"id":"https://openalex.org/C58013763","wikidata":"https://www.wikidata.org/wiki/Q5754574","display_name":"High-level synthesis","level":3,"score":0.2653000056743622},{"id":"https://openalex.org/C186370098","wikidata":"https://www.wikidata.org/wiki/Q442787","display_name":"Energy (signal processing)","level":2,"score":0.2565000057220459},{"id":"https://openalex.org/C56372850","wikidata":"https://www.wikidata.org/wiki/Q1050404","display_name":"Sparse matrix","level":3,"score":0.2549000084400177},{"id":"https://openalex.org/C118021083","wikidata":"https://www.wikidata.org/wiki/Q610398","display_name":"System on a chip","level":2,"score":0.2508000135421753}],"mesh":[],"locations_count":1,"locations":[{"id":"doi:10.1109/access.2025.3625771","is_oa":true,"landing_page_url":"https://doi.org/10.1109/access.2025.3625771","pdf_url":null,"source":{"id":"https://openalex.org/S2485537415","display_name":"IEEE Access","issn_l":"2169-3536","issn":["2169-3536"],"is_oa":true,"is_in_doaj":true,"is_core":true,"host_organization":"https://openalex.org/P4310319808","host_organization_name":"Institute of Electrical and Electronics Engineers","host_organization_lineage":["https://openalex.org/P4310319808"],"host_organization_lineage_names":["Institute of Electrical and Electronics Engineers"],"type":"journal"},"license":"cc-by","license_id":"https://openalex.org/licenses/cc-by","version":"publishedVersion","is_accepted":true,"is_published":true,"raw_source_name":"IEEE Access","raw_type":"journal-article"}],"best_oa_location":{"id":"doi:10.1109/access.2025.3625771","is_oa":true,"landing_page_url":"https://doi.org/10.1109/access.2025.3625771","pdf_url":null,"source":{"id":"https://openalex.org/S2485537415","display_name":"IEEE Access","issn_l":"2169-3536","issn":["2169-3536"],"is_oa":true,"is_in_doaj":true,"is_core":true,"host_organization":"https://openalex.org/P4310319808","host_organization_name":"Institute of Electrical and Electronics Engineers","host_organization_lineage":["https://openalex.org/P4310319808"],"host_organization_lineage_names":["Institute of Electrical and Electronics Engineers"],"type":"journal"},"license":"cc-by","license_id":"https://openalex.org/licenses/cc-by","version":"publishedVersion","is_accepted":true,"is_published":true,"raw_source_name":"IEEE Access","raw_type":"journal-article"},"sustainable_development_goals":[],"awards":[],"funders":[{"id":"https://openalex.org/F4320307910","display_name":"Booz Allen Hamilton","ror":"https://ror.org/051rcp357"}],"has_content":{"grobid_xml":false,"pdf":false},"content_urls":null,"referenced_works_count":23,"referenced_works":["https://openalex.org/W2946609015","https://openalex.org/W2981751377","https://openalex.org/W2982041622","https://openalex.org/W3047848469","https://openalex.org/W3115388607","https://openalex.org/W3130240120","https://openalex.org/W3162542754","https://openalex.org/W3199934250","https://openalex.org/W3206837665","https://openalex.org/W4308479898","https://openalex.org/W4391125526","https://openalex.org/W4393141090","https://openalex.org/W4393146966","https://openalex.org/W4393147284","https://openalex.org/W4393578753","https://openalex.org/W4393973515","https://openalex.org/W4399487406","https://openalex.org/W4400234428","https://openalex.org/W4401567938","https://openalex.org/W4402350363","https://openalex.org/W4402670692","https://openalex.org/W4406650295","https://openalex.org/W4409282814"],"related_works":[],"abstract_inverted_index":{"While":[0],"Large":[1],"Language":[2],"Models":[3],"(LLMs)":[4],"have":[5],"demonstrated":[6],"remarkable":[7],"performance,":[8],"their":[9,18,84],"deployment":[10,191],"on":[11,83,86,197],"resource-constrained":[12],"edge":[13,198],"devices":[14],"is":[15],"hindered":[16],"by":[17],"immense":[19],"size.":[20],"Existing":[21],"quantization":[22,49,62,104],"methods":[23],"offer":[24],"a":[25,41,46,52,100,124,129,152,166],"promising":[26],"solution":[27],"but":[28],"struggle":[29],"to":[30,56,70,80,91,179],"achieve":[31],"sub-4-bit":[32],"precision":[33],"without":[34],"significant":[35],"performance":[36],"degradation.":[37],"This":[38,135,159],"work":[39],"introduces":[40],"holistic":[42],"framework":[43],"that":[44,106],"pairs":[45],"novel":[47],"mixed-precision":[48],"algorithm":[50],"with":[51],"custom":[53],"hardware":[54,126,164],"accelerator":[55,127],"address":[57],"these":[58,95,117],"challenges.":[59],"Our":[60],"post-training":[61],"method":[63],"leverages":[64],"an":[65,143],"in-depth":[66],"analysis":[67],"of":[68,113,161,192],"sensitivity":[69],"allocate":[71],"varying":[72],"bit":[73],"precisions":[74],"(FP16,":[75],"INT2,":[76],"INT3,":[77],"and":[78,151,163,174,184,194],"INT4)":[79],"weights":[81,150],"based":[82],"impact":[85],"model":[87,170,181],"performance.":[88],"In":[89],"addition,":[90],"enhance":[92],"accuracy":[93],"at":[94],"low":[96],"bit-widths,":[97],"we":[98,122],"introduce":[99],"scaled":[101],"power-of-two":[102],"logarithmic":[103,145],"scheme":[105],"more":[107],"effectively":[108],"captures":[109],"the":[110,190],"weight":[111],"distributions":[112],"LLMs.":[114],"To":[115],"realize":[116],"algorithmic":[118],"gains":[119],"in":[120,177],"practice,":[121],"co-design":[123,160],"dedicated":[125],"featuring":[128],"Dual-Mode":[130],"Matrix":[131],"Multiplication":[132],"Unit":[133],"(DM-MMU).":[134],"dual-mode":[136],"matrix":[137],"multiplication":[138],"unit":[139],"dynamically":[140],"switches":[141],"between":[142,169],"efficient":[144],"datapath":[146,155],"for":[147,156],"log":[148],"quantized":[149],"full-precision":[153],"FP16":[154],"other":[157],"calculations.":[158],"software":[162],"achieves":[165],"compelling":[167],"trade-off":[168],"accuracy,":[171],"memory":[172],"footprint,":[173],"computational":[175],"efficiency\u2014resulting":[176],"up":[178],"11.8\u00d7":[180],"size":[182],"reduction":[183],"1.82\u00d7":[185],"higher":[186],"energy":[187],"efficiency,":[188],"enabling":[189],"powerful":[193],"lightweight":[195],"LLMs":[196],"devices.":[199]},"counts_by_year":[],"updated_date":"2025-12-16T23:43:54.943958","created_date":"2025-10-28T00:00:00"}
