{"id":"https://openalex.org/W4415821472","doi":"https://doi.org/10.1109/tcasai.2025.3628250","title":"Softprox: A Post-Finetuning Methodology to Mitigate Softmax Bottlenecks in Transformer Workloads","display_name":"Softprox: A Post-Finetuning Methodology to Mitigate Softmax Bottlenecks in Transformer Workloads","publication_year":2025,"publication_date":"2025-11-03","ids":{"openalex":"https://openalex.org/W4415821472","doi":"https://doi.org/10.1109/tcasai.2025.3628250"},"language":null,"primary_location":{"id":"doi:10.1109/tcasai.2025.3628250","is_oa":false,"landing_page_url":"https://doi.org/10.1109/tcasai.2025.3628250","pdf_url":null,"source":{"id":"https://openalex.org/S4404675360","display_name":"IEEE transactions on circuits and systems for artificial intelligence.","issn_l":"2996-6647","issn":["2996-6647"],"is_oa":false,"is_in_doaj":false,"is_core":false,"host_organization":null,"host_organization_name":null,"host_organization_lineage":[],"host_organization_lineage_names":[],"type":"journal"},"license":null,"license_id":null,"version":"publishedVersion","is_accepted":true,"is_published":true,"raw_source_name":"IEEE Transactions on Circuits and Systems for Artificial Intelligence","raw_type":"journal-article"},"type":"article","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/A5001044510","display_name":"Aradhana Mohan Parvathy","orcid":null},"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":"Aradhana Mohan Parvathy","raw_affiliation_strings":["Purdue University, West Lafayette, IN, USA"],"raw_orcid":"https://orcid.org/0000-0001-6961-3540","affiliations":[{"raw_affiliation_string":"Purdue University, West Lafayette, IN, USA","institution_ids":["https://openalex.org/I219193219"]}]},{"author_position":"middle","author":{"id":"https://openalex.org/A5083686495","display_name":"Sourjya Roy","orcid":"https://orcid.org/0000-0002-6864-7738"},"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":"Sourjya Roy","raw_affiliation_strings":["Purdue University, West Lafayette, IN, USA"],"raw_orcid":"https://orcid.org/0000-0002-6864-7738","affiliations":[{"raw_affiliation_string":"Purdue University, West Lafayette, IN, USA","institution_ids":["https://openalex.org/I219193219"]}]},{"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"]},{"id":"https://openalex.org/I4210158342","display_name":"Intel (United Kingdom)","ror":"https://ror.org/058cxws58","country_code":"GB","type":"company","lineage":["https://openalex.org/I1343180700","https://openalex.org/I4210158342"]}],"countries":["GB","US"],"is_corresponding":false,"raw_author_name":"Soumendu Kumar Ghosh","raw_affiliation_strings":["Intel Corporation, Santa Clara, CA, USA","Intel Corporation"],"raw_orcid":"https://orcid.org/0000-0001-6776-1427","affiliations":[{"raw_affiliation_string":"Intel Corporation, Santa Clara, CA, USA","institution_ids":["https://openalex.org/I1343180700"]},{"raw_affiliation_string":"Intel Corporation","institution_ids":["https://openalex.org/I4210158342"]}]},{"author_position":"middle","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"]},{"id":"https://openalex.org/I4210158342","display_name":"Intel (United Kingdom)","ror":"https://ror.org/058cxws58","country_code":"GB","type":"company","lineage":["https://openalex.org/I1343180700","https://openalex.org/I4210158342"]}],"countries":["GB","US"],"is_corresponding":false,"raw_author_name":"Arnab Raha","raw_affiliation_strings":["Intel Corporation, Santa Clara, CA, USA","Intel Corporation"],"raw_orcid":"https://orcid.org/0000-0002-8848-1069","affiliations":[{"raw_affiliation_string":"Intel Corporation, Santa Clara, CA, USA","institution_ids":["https://openalex.org/I1343180700"]},{"raw_affiliation_string":"Intel Corporation","institution_ids":["https://openalex.org/I4210158342"]}]},{"author_position":"middle","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"]},{"id":"https://openalex.org/I4210158342","display_name":"Intel (United Kingdom)","ror":"https://ror.org/058cxws58","country_code":"GB","type":"company","lineage":["https://openalex.org/I1343180700","https://openalex.org/I4210158342"]}],"countries":["GB","US"],"is_corresponding":false,"raw_author_name":"Deepak Mathaikutty","raw_affiliation_strings":["Intel Corporation, Santa Clara, CA, USA","Intel Corporation"],"raw_orcid":"https://orcid.org/0009-0004-3768-1337","affiliations":[{"raw_affiliation_string":"Intel Corporation, Santa Clara, CA, USA","institution_ids":["https://openalex.org/I1343180700"]},{"raw_affiliation_string":"Intel Corporation","institution_ids":["https://openalex.org/I4210158342"]}]},{"author_position":"last","author":{"id":"https://openalex.org/A5065766721","display_name":"Anand Raghunathan","orcid":"https://orcid.org/0000-0002-4624-564X"},"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":"Anand Raghunathan","raw_affiliation_strings":["Purdue University, West Lafayette, IN, USA"],"raw_orcid":"https://orcid.org/0000-0002-4624-564X","affiliations":[{"raw_affiliation_string":"Purdue University, West Lafayette, IN, USA","institution_ids":["https://openalex.org/I219193219"]}]}],"institutions":[],"countries_distinct_count":2,"institutions_distinct_count":6,"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":0.27913199,"is_in_top_1_percent":false,"is_in_top_10_percent":false},"cited_by_percentile_year":null,"biblio":{"volume":"3","issue":"2","first_page":"90","last_page":"100"},"is_retracted":false,"is_paratext":false,"is_xpac":false,"primary_topic":{"id":"https://openalex.org/T10036","display_name":"Advanced Neural Network Applications","score":0.6434999704360962,"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":0.6434999704360962,"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/T10775","display_name":"Generative Adversarial Networks and Image Synthesis","score":0.05730000138282776,"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/T11714","display_name":"Multimodal Machine Learning Applications","score":0.020999999716877937,"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"}}],"keywords":[{"id":"https://openalex.org/keywords/softmax-function","display_name":"Softmax function","score":0.977400004863739},{"id":"https://openalex.org/keywords/exponentiation","display_name":"Exponentiation","score":0.7239999771118164},{"id":"https://openalex.org/keywords/computation","display_name":"Computation","score":0.6043000221252441},{"id":"https://openalex.org/keywords/computational-complexity-theory","display_name":"Computational complexity theory","score":0.5443000197410583},{"id":"https://openalex.org/keywords/transformer","display_name":"Transformer","score":0.5159000158309937},{"id":"https://openalex.org/keywords/artificial-neural-network","display_name":"Artificial neural network","score":0.44929999113082886},{"id":"https://openalex.org/keywords/computational-resource","display_name":"Computational resource","score":0.37070000171661377}],"concepts":[{"id":"https://openalex.org/C188441871","wikidata":"https://www.wikidata.org/wiki/Q7554146","display_name":"Softmax function","level":3,"score":0.977400004863739},{"id":"https://openalex.org/C81539297","wikidata":"https://www.wikidata.org/wiki/Q33456","display_name":"Exponentiation","level":2,"score":0.7239999771118164},{"id":"https://openalex.org/C41008148","wikidata":"https://www.wikidata.org/wiki/Q21198","display_name":"Computer science","level":0,"score":0.7232000231742859},{"id":"https://openalex.org/C45374587","wikidata":"https://www.wikidata.org/wiki/Q12525525","display_name":"Computation","level":2,"score":0.6043000221252441},{"id":"https://openalex.org/C179799912","wikidata":"https://www.wikidata.org/wiki/Q205084","display_name":"Computational complexity theory","level":2,"score":0.5443000197410583},{"id":"https://openalex.org/C66322947","wikidata":"https://www.wikidata.org/wiki/Q11658","display_name":"Transformer","level":3,"score":0.5159000158309937},{"id":"https://openalex.org/C50644808","wikidata":"https://www.wikidata.org/wiki/Q192776","display_name":"Artificial neural network","level":2,"score":0.44929999113082886},{"id":"https://openalex.org/C154945302","wikidata":"https://www.wikidata.org/wiki/Q11660","display_name":"Artificial intelligence","level":1,"score":0.4440000057220459},{"id":"https://openalex.org/C127964446","wikidata":"https://www.wikidata.org/wiki/Q1092142","display_name":"Computational resource","level":3,"score":0.37070000171661377},{"id":"https://openalex.org/C55526617","wikidata":"https://www.wikidata.org/wiki/Q719375","display_name":"Operand","level":2,"score":0.3515999913215637},{"id":"https://openalex.org/C206588197","wikidata":"https://www.wikidata.org/wiki/Q846574","display_name":"Reuse","level":2,"score":0.32899999618530273},{"id":"https://openalex.org/C11413529","wikidata":"https://www.wikidata.org/wiki/Q8366","display_name":"Algorithm","level":1,"score":0.3278000056743622},{"id":"https://openalex.org/C2780513914","wikidata":"https://www.wikidata.org/wiki/Q18210350","display_name":"Bottleneck","level":2,"score":0.3208000063896179},{"id":"https://openalex.org/C81363708","wikidata":"https://www.wikidata.org/wiki/Q17084460","display_name":"Convolutional neural network","level":2,"score":0.3093000054359436},{"id":"https://openalex.org/C119857082","wikidata":"https://www.wikidata.org/wiki/Q2539","display_name":"Machine learning","level":1,"score":0.30309998989105225},{"id":"https://openalex.org/C108583219","wikidata":"https://www.wikidata.org/wiki/Q197536","display_name":"Deep learning","level":2,"score":0.2888999879360199},{"id":"https://openalex.org/C66024118","wikidata":"https://www.wikidata.org/wiki/Q1122506","display_name":"Computational model","level":2,"score":0.2858000099658966},{"id":"https://openalex.org/C113775141","wikidata":"https://www.wikidata.org/wiki/Q428691","display_name":"Computer engineering","level":1,"score":0.28110000491142273},{"id":"https://openalex.org/C34736171","wikidata":"https://www.wikidata.org/wiki/Q918333","display_name":"Preprocessor","level":2,"score":0.2784999907016754},{"id":"https://openalex.org/C80444323","wikidata":"https://www.wikidata.org/wiki/Q2878974","display_name":"Theoretical computer science","level":1,"score":0.26570001244544983}],"mesh":[],"locations_count":1,"locations":[{"id":"doi:10.1109/tcasai.2025.3628250","is_oa":false,"landing_page_url":"https://doi.org/10.1109/tcasai.2025.3628250","pdf_url":null,"source":{"id":"https://openalex.org/S4404675360","display_name":"IEEE transactions on circuits and systems for artificial intelligence.","issn_l":"2996-6647","issn":["2996-6647"],"is_oa":false,"is_in_doaj":false,"is_core":false,"host_organization":null,"host_organization_name":null,"host_organization_lineage":[],"host_organization_lineage_names":[],"type":"journal"},"license":null,"license_id":null,"version":"publishedVersion","is_accepted":true,"is_published":true,"raw_source_name":"IEEE Transactions on Circuits and Systems for Artificial Intelligence","raw_type":"journal-article"}],"best_oa_location":null,"sustainable_development_goals":[],"awards":[],"funders":[{"id":"https://openalex.org/F4320332180","display_name":"Defense Advanced Research Projects Agency","ror":"https://ror.org/02caytj08"}],"has_content":{"pdf":false,"grobid_xml":false},"content_urls":null,"referenced_works_count":13,"referenced_works":["https://openalex.org/W2108598243","https://openalex.org/W2896457183","https://openalex.org/W2923014074","https://openalex.org/W3094502228","https://openalex.org/W3172942063","https://openalex.org/W3197580070","https://openalex.org/W3211525823","https://openalex.org/W4293025109","https://openalex.org/W4385245566","https://openalex.org/W4392867133","https://openalex.org/W4394625669","https://openalex.org/W4413442950","https://openalex.org/W7133227460"],"related_works":[],"abstract_inverted_index":{"Transformer":[0,244],"neural":[1,208],"networks":[2],"have":[3],"emerged":[4],"as":[5,95],"the":[6,19,46,64,70,74,80,97,122,181,190,212,226,268],"state-of-the-art":[7],"in":[8,50,167],"AI":[9],"across":[10],"text,":[11,247],"audio,":[12,248],"image,":[13],"and":[14,76,90,160,195,220,249],"video":[15],"processing":[16],"tasks.":[17],"However,":[18],"attention":[20,47,175,183],"mechanism":[21,178],"that":[22,30,109,144,179,197],"is":[23,49,173],"core":[24],"to":[25,104,120,163,210],"Transformers":[26],"has":[27],"a":[28,39,52,141,206,230,235,241,262],"complexity":[29,72],"scales":[31],"quadratically":[32],"with":[33,234,261],"input":[34],"sequence":[35],"length,":[36],"making":[37],"it":[38],"computational":[40,65,123],"bottleneck.":[41,57],"The":[42],"Softmax":[43,126,147,165,187,213],"operation":[44],"within":[45,85],"computation":[48],"turn":[51],"significant":[53],"reason":[54],"for":[55,63,112,229,246],"this":[56],"We":[58],"identify":[59],"three":[60,154],"primary":[61],"reasons":[62],"challenge":[66],"of":[67,73,125,243,259,265],"Softmax:":[68],"(1)":[69],"intrinsic":[71],"exponentiation":[75,219],"division":[77,221],"operations,":[78],"(2)":[79],"data":[81],"dependencies":[82],"among":[83],"operations":[84,93],"Softmax,":[86],"which":[87,99],"limit":[88],"parallelism,":[89],"(3)":[91],"reduction":[92],"such":[94],"finding":[96],"maximum,":[98],"do":[100],"not":[101],"map":[102],"well":[103],"NPUs":[105],"or":[106,131,133],"MAC":[107],"Arrays":[108],"are":[110],"designed":[111],"matrix":[113],"multiply":[114],"operations.":[115,222],"<p":[116],"xmlns:mml=\"http://www.w3.org/1998/Math/MathML\"":[117],"xmlns:xlink=\"http://www.w3.org/1999/xlink\">Existing":[118],"approaches":[119],"improve":[121],"efficiency":[124],"often":[127],"require":[128],"extensive":[129],"pretraining":[130],"fine-tuning,":[132],"specialized":[134],"hardware.":[135,151],"In":[136],"contrast,":[137],"we":[138],"propose":[139],"<monospace>Softprox</monospace>,":[140],"post-fine-tuning":[142],"technique":[143],"can":[145],"accelerate":[146],"computations":[148,166,184],"on":[149,217,267],"commodity":[150],"<monospace>Softprox</monospace>":[152,252],"utilizes":[153],"complementary":[155],"approximations":[156],"\u2012":[157,162],"AttSkip,":[158],"SoftNet,":[159],"Stat-Max":[161],"optimize":[164],"any":[168],"given":[169],"finetuned":[170],"network.":[171],"AttSkip":[172,200],"an":[174,254],"score":[176],"reuse":[177],"minimizes":[180],"required":[182],"by":[185],"reusing":[186],"scores":[188],"from":[189],"previous":[191],"layer.":[192],"For":[193],"layers":[194],"heads":[196],"cannot":[198],"adopt":[199],"without":[201],"accuracy":[202],"degradation,":[203],"SoftNet":[204],"employs":[205],"low-complexity":[207],"network":[209],"approximate":[211],"operation,":[214],"reducing":[215],"reliance":[216],"complex":[218],"Finally,":[223],"StatMax":[224],"replaces":[225],"dynamic":[227],"search":[228],"vector\u2019s":[231],"maximum":[232,238,263],"element":[233],"statistically":[236],"precomputed":[237],"value.":[239],"Across":[240],"suite":[242],"models":[245],"vision":[250],"tasks,":[251],"demonstrates":[253],"average":[255],"runtime":[256],"inference":[257],"speedup":[258,264],"21.72%,":[260],"40.22%":[266],"Intel<sup>\u00ae</sup>":[269],"Core<sup><small>\u2122</small></sup>":[270],"Ultra":[271],"Series":[272],"1":[273],"NPU.":[274]},"counts_by_year":[],"updated_date":"2026-06-11T09:08:48.828518","created_date":"2025-11-03T00:00:00"}