{"id":"https://openalex.org/W7126244407","doi":"https://doi.org/10.1109/lca.2026.3659512","title":"De-Quantization Penalties for Interactive LLM Inference on Prosumer GPUs","display_name":"De-Quantization Penalties for Interactive LLM Inference on Prosumer GPUs","publication_year":2026,"publication_date":"2026-01-01","ids":{"openalex":"https://openalex.org/W7126244407","doi":"https://doi.org/10.1109/lca.2026.3659512"},"language":null,"primary_location":{"id":"doi:10.1109/lca.2026.3659512","is_oa":false,"landing_page_url":"https://doi.org/10.1109/lca.2026.3659512","pdf_url":null,"source":{"id":"https://openalex.org/S17643076","display_name":"IEEE Computer Architecture Letters","issn_l":"1556-6056","issn":["1556-6056","1556-6064","2473-2575"],"is_oa":false,"is_in_doaj":false,"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":null,"license_id":null,"version":"publishedVersion","is_accepted":true,"is_published":true,"raw_source_name":"IEEE Computer Architecture Letters","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/A5011729764","display_name":"Junaid Ahmad Khan","orcid":"https://orcid.org/0000-0001-7211-4646"},"institutions":[{"id":"https://openalex.org/I118373667","display_name":"Seoul National University of Science and Technology","ror":"https://ror.org/00chfja07","country_code":"KR","type":"education","lineage":["https://openalex.org/I118373667"]}],"countries":["KR"],"is_corresponding":true,"raw_author_name":"Junaid Ahmad Khan","raw_affiliation_strings":["Department of Electrical and Information Engineering, Seoul National University of Science and Technology, Seoul, South Korea"],"raw_orcid":"https://orcid.org/0000-0001-7211-4646","affiliations":[{"raw_affiliation_string":"Department of Electrical and Information Engineering, Seoul National University of Science and Technology, Seoul, South Korea","institution_ids":["https://openalex.org/I118373667"]}]}],"institutions":[],"countries_distinct_count":1,"institutions_distinct_count":1,"corresponding_author_ids":["https://openalex.org/A5011729764"],"corresponding_institution_ids":["https://openalex.org/I118373667"],"apc_list":null,"apc_paid":null,"fwci":0.0,"has_fulltext":false,"cited_by_count":0,"citation_normalized_percentile":{"value":0.18029876,"is_in_top_1_percent":false,"is_in_top_10_percent":false},"cited_by_percentile_year":null,"biblio":{"volume":"25","issue":"1","first_page":"45","last_page":"48"},"is_retracted":false,"is_paratext":false,"is_xpac":false,"primary_topic":{"id":"https://openalex.org/T10036","display_name":"Advanced Neural Network Applications","score":0.35580000281333923,"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":0.35580000281333923,"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/T10181","display_name":"Natural Language Processing Techniques","score":0.1362999975681305,"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":0.053599998354911804,"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/inference","display_name":"Inference","score":0.6660000085830688},{"id":"https://openalex.org/keywords/quantization","display_name":"Quantization (signal processing)","score":0.5605999827384949},{"id":"https://openalex.org/keywords/decoding-methods","display_name":"Decoding methods","score":0.5307000279426575},{"id":"https://openalex.org/keywords/latency","display_name":"Latency (audio)","score":0.4876999855041504},{"id":"https://openalex.org/keywords/prosumer","display_name":"Prosumer","score":0.4657000005245209},{"id":"https://openalex.org/keywords/byte","display_name":"Byte","score":0.42809998989105225},{"id":"https://openalex.org/keywords/kernel","display_name":"Kernel (algebra)","score":0.35839998722076416},{"id":"https://openalex.org/keywords/throughput","display_name":"Throughput","score":0.33640000224113464}],"concepts":[{"id":"https://openalex.org/C41008148","wikidata":"https://www.wikidata.org/wiki/Q21198","display_name":"Computer science","level":0,"score":0.8331999778747559},{"id":"https://openalex.org/C2776214188","wikidata":"https://www.wikidata.org/wiki/Q408386","display_name":"Inference","level":2,"score":0.6660000085830688},{"id":"https://openalex.org/C28855332","wikidata":"https://www.wikidata.org/wiki/Q198099","display_name":"Quantization (signal processing)","level":2,"score":0.5605999827384949},{"id":"https://openalex.org/C57273362","wikidata":"https://www.wikidata.org/wiki/Q576722","display_name":"Decoding methods","level":2,"score":0.5307000279426575},{"id":"https://openalex.org/C82876162","wikidata":"https://www.wikidata.org/wiki/Q17096504","display_name":"Latency (audio)","level":2,"score":0.4876999855041504},{"id":"https://openalex.org/C2779939747","wikidata":"https://www.wikidata.org/wiki/Q1145220","display_name":"Prosumer","level":3,"score":0.4657000005245209},{"id":"https://openalex.org/C43364308","wikidata":"https://www.wikidata.org/wiki/Q8799","display_name":"Byte","level":2,"score":0.42809998989105225},{"id":"https://openalex.org/C173608175","wikidata":"https://www.wikidata.org/wiki/Q232661","display_name":"Parallel computing","level":1,"score":0.3808000087738037},{"id":"https://openalex.org/C74193536","wikidata":"https://www.wikidata.org/wiki/Q574844","display_name":"Kernel (algebra)","level":2,"score":0.35839998722076416},{"id":"https://openalex.org/C157764524","wikidata":"https://www.wikidata.org/wiki/Q1383412","display_name":"Throughput","level":3,"score":0.33640000224113464},{"id":"https://openalex.org/C3115412","wikidata":"https://www.wikidata.org/wiki/Q1194708","display_name":"Bitmap","level":2,"score":0.33239999413490295},{"id":"https://openalex.org/C46686674","wikidata":"https://www.wikidata.org/wiki/Q466303","display_name":"Boosting (machine learning)","level":2,"score":0.32409998774528503},{"id":"https://openalex.org/C165696696","wikidata":"https://www.wikidata.org/wiki/Q11287","display_name":"Exploit","level":2,"score":0.32030001282691956},{"id":"https://openalex.org/C137293760","wikidata":"https://www.wikidata.org/wiki/Q3621696","display_name":"Language model","level":2,"score":0.31459999084472656},{"id":"https://openalex.org/C115874739","wikidata":"https://www.wikidata.org/wiki/Q825377","display_name":"Critical path method","level":2,"score":0.3109999895095825},{"id":"https://openalex.org/C2777735758","wikidata":"https://www.wikidata.org/wiki/Q817765","display_name":"Path (computing)","level":2,"score":0.3050999939441681},{"id":"https://openalex.org/C155281189","wikidata":"https://www.wikidata.org/wiki/Q3518150","display_name":"Tensor (intrinsic definition)","level":2,"score":0.304500013589859},{"id":"https://openalex.org/C11413529","wikidata":"https://www.wikidata.org/wiki/Q8366","display_name":"Algorithm","level":1,"score":0.299699991941452},{"id":"https://openalex.org/C111335779","wikidata":"https://www.wikidata.org/wiki/Q3454686","display_name":"Reduction (mathematics)","level":2,"score":0.2888999879360199},{"id":"https://openalex.org/C86111242","wikidata":"https://www.wikidata.org/wiki/Q859595","display_name":"Coprocessor","level":2,"score":0.2849999964237213},{"id":"https://openalex.org/C80444323","wikidata":"https://www.wikidata.org/wiki/Q2878974","display_name":"Theoretical computer science","level":1,"score":0.27880001068115234},{"id":"https://openalex.org/C186967261","wikidata":"https://www.wikidata.org/wiki/Q5082128","display_name":"Mobile device","level":2,"score":0.2703999876976013},{"id":"https://openalex.org/C113775141","wikidata":"https://www.wikidata.org/wiki/Q428691","display_name":"Computer engineering","level":1,"score":0.26919999718666077}],"mesh":[],"locations_count":1,"locations":[{"id":"doi:10.1109/lca.2026.3659512","is_oa":false,"landing_page_url":"https://doi.org/10.1109/lca.2026.3659512","pdf_url":null,"source":{"id":"https://openalex.org/S17643076","display_name":"IEEE Computer Architecture Letters","issn_l":"1556-6056","issn":["1556-6056","1556-6064","2473-2575"],"is_oa":false,"is_in_doaj":false,"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":null,"license_id":null,"version":"publishedVersion","is_accepted":true,"is_published":true,"raw_source_name":"IEEE Computer Architecture Letters","raw_type":"journal-article"}],"best_oa_location":null,"sustainable_development_goals":[{"score":0.9213564991950989,"display_name":"Affordable and clean energy","id":"https://metadata.un.org/sdg/7"}],"awards":[],"funders":[],"has_content":{"grobid_xml":false,"pdf":false},"content_urls":null,"referenced_works_count":0,"referenced_works":[],"related_works":[],"abstract_inverted_index":{"Post-training":[0],"4-bit":[1,29,73,127,159],"quantization":[2,160],"is":[3,21,95,139,161],"often":[4],"treated":[5],"as":[6],"the":[7,136,173],"default":[8],"path":[9],"for":[10,177],"running":[11],"large":[12],"language":[13],"models":[14,59,74],"(LLMs)":[15],"on":[16,44,151],"prosumer":[17,152],"GPUs:":[18],"if":[19],"decoding":[20,94],"memory-bandwidth":[22],"bound,":[23],"shrinking":[24],"weights":[25],"from":[26],"FP16":[27,69,102,171],"to":[28,116],"should":[30],"cut":[31],"memory":[32],"traffic":[33],"and":[34,37,75,110],"improve":[35],"latency":[36],"energy":[38],"efficiency.":[39],"We":[40,148],"revisit":[41],"this":[42],"assumption":[43],"an":[45],"RTX":[46],"3090":[47],"(Ampere)":[48],"that":[49,135,150],"lacks":[50],"native":[51,68,155],"INT4":[52,112,156],"tensor":[53,157],"support.":[54],"For":[55],"1\u20138":[56],"billion":[57],"parameter":[58],"(TinyLlama,":[60],"Qwen-2.5,":[61],"Mistral,":[62],"Llama":[63],"3.1,":[64],"DeepSeek-R1-8B),":[65],"we":[66],"compare":[67],"inference":[70],"against":[71],"AutoGPTQ":[72],"GGUF":[76,124],"kernels":[77],"in":[78,107],"<monospace":[79],"xmlns:mml=\"http://www.w3.org/1998/Math/MathML\"":[80,91,98,118],"xmlns:xlink=\"http://www.w3.org/1999/xlink\">llama.cpp</monospace>.":[81],"On":[82],"a":[83,104],"standard":[84],"Transformers+AutoGPTQ":[85],"stack,":[86],"interactive":[87,178],"batch":[88],"size":[89],"<inline-formula":[90],"xmlns:xlink=\"http://www.w3.org/1999/xlink\"><tex-math":[92],"notation=\"LaTeX\">$B=1$</tex-math></inline-formula>":[93],"still":[96],"<bold":[97,117],"xmlns:xlink=\"http://www.w3.org/1999/xlink\">1.3\u20132.2\u00d7":[99],"slower</b>":[100],"than":[101,145],"despite":[103],"2.4\u00d7":[105],"reduction":[106],"VRAM":[108],"usage,":[109],"some":[111],"configurations":[113],"are":[114],"up":[115],"xmlns:xlink=\"http://www.w3.org/1999/xlink\">2.4\u00d7":[119],"less":[120],"energy-efficient</b>.":[121],"An":[122],"optimized":[123],"backend":[125],"improves":[126],"TinyLlama":[128],"throughput":[129],"by":[130,141],"1.65\u00d7":[131],"over":[132],"GPTQ,":[133],"indicating":[134],"de-quantization":[137],"penalty":[138],"dominated":[140],"kernel":[142],"design":[143],"rather":[144],"hardware":[146],"limits.":[147],"conclude":[149],"GPUs":[153],"without":[154],"cores,":[158],"only":[162],"attractive":[163],"when":[164],"paired":[165],"with":[166],"mature":[167],"low-bit":[168],"kernels;":[169],"otherwise,":[170],"remains":[172],"more":[174],"robust":[175],"choice":[176],"workloads.":[179]},"counts_by_year":[],"updated_date":"2026-02-21T06:11:54.161237","created_date":"2026-02-01T00:00:00"}