{"id":"https://openalex.org/W7125909771","doi":"https://doi.org/10.1145/3774934.3786423","title":"High-Throughput Non-uniformly Quantized 3-bit LLM Inference","display_name":"High-Throughput Non-uniformly Quantized 3-bit LLM Inference","publication_year":2026,"publication_date":"2026-01-28","ids":{"openalex":"https://openalex.org/W7125909771","doi":"https://doi.org/10.1145/3774934.3786423"},"language":null,"primary_location":{"id":"doi:10.1145/3774934.3786423","is_oa":false,"landing_page_url":"https://doi.org/10.1145/3774934.3786423","pdf_url":null,"source":null,"license":null,"license_id":null,"version":"publishedVersion","is_accepted":true,"is_published":true,"raw_source_name":"Proceedings of the 31st ACM SIGPLAN Annual Symposium on Principles and Practice of Parallel Programming","raw_type":"proceedings-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/A5050067045","display_name":"YuAng Chen","orcid":"https://orcid.org/0000-0002-3392-8388"},"institutions":[{"id":"https://openalex.org/I177725633","display_name":"Chinese University of Hong Kong","ror":"https://ror.org/00t33hh48","country_code":"HK","type":"education","lineage":["https://openalex.org/I177725633"]}],"countries":["HK"],"is_corresponding":true,"raw_author_name":"YuAng Chen","raw_affiliation_strings":["Chinese University of Hong Kong, Hong Kong, China"],"affiliations":[{"raw_affiliation_string":"Chinese University of Hong Kong, Hong Kong, China","institution_ids":["https://openalex.org/I177725633"]}]},{"author_position":"middle","author":{"id":null,"display_name":"Wenqi Zeng","orcid":"https://orcid.org/0000-0002-9770-6522"},"institutions":[{"id":"https://openalex.org/I200769079","display_name":"Hong Kong University of Science and Technology","ror":"https://ror.org/00q4vv597","country_code":"HK","type":"education","lineage":["https://openalex.org/I200769079"]}],"countries":["HK"],"is_corresponding":false,"raw_author_name":"Wenqi Zeng","raw_affiliation_strings":["Hong Kong University of Science and Technology, Hong Kong, China"],"affiliations":[{"raw_affiliation_string":"Hong Kong University of Science and Technology, Hong Kong, China","institution_ids":["https://openalex.org/I200769079"]}]},{"author_position":"last","author":{"id":"https://openalex.org/A5124121088","display_name":"Jeffrey Xu Yu","orcid":null},"institutions":[{"id":"https://openalex.org/I200769079","display_name":"Hong Kong University of Science and Technology","ror":"https://ror.org/00q4vv597","country_code":"HK","type":"education","lineage":["https://openalex.org/I200769079"]}],"countries":["HK"],"is_corresponding":false,"raw_author_name":"Jeffrey Xu Yu","raw_affiliation_strings":["Hong Kong University of Science and Technology (Guangzhou), Hong Kong, China"],"affiliations":[{"raw_affiliation_string":"Hong Kong University of Science and Technology (Guangzhou), Hong Kong, China","institution_ids":["https://openalex.org/I200769079"]}]}],"institutions":[],"countries_distinct_count":1,"institutions_distinct_count":3,"corresponding_author_ids":["https://openalex.org/A5050067045"],"corresponding_institution_ids":["https://openalex.org/I177725633"],"apc_list":null,"apc_paid":null,"fwci":0.0,"has_fulltext":false,"cited_by_count":0,"citation_normalized_percentile":{"value":0.19476281,"is_in_top_1_percent":false,"is_in_top_10_percent":false},"cited_by_percentile_year":null,"biblio":{"volume":null,"issue":null,"first_page":"288","last_page":"300"},"is_retracted":false,"is_paratext":false,"is_xpac":false,"primary_topic":{"id":"https://openalex.org/T10036","display_name":"Advanced Neural Network Applications","score":0.23960000276565552,"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":0.23960000276565552,"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/T10054","display_name":"Parallel Computing and Optimization Techniques","score":0.2214999943971634,"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":0.10750000178813934,"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"}}],"keywords":[{"id":"https://openalex.org/keywords/quantization","display_name":"Quantization (signal processing)","score":0.5940999984741211},{"id":"https://openalex.org/keywords/pipeline","display_name":"Pipeline (software)","score":0.5824000239372253},{"id":"https://openalex.org/keywords/key","display_name":"Key (lock)","score":0.5012000203132629},{"id":"https://openalex.org/keywords/shuffling","display_name":"Shuffling","score":0.4966000020503998},{"id":"https://openalex.org/keywords/scheme","display_name":"Scheme (mathematics)","score":0.43970000743865967},{"id":"https://openalex.org/keywords/inference","display_name":"Inference","score":0.4359999895095825},{"id":"https://openalex.org/keywords/cuda","display_name":"CUDA","score":0.4255000054836273},{"id":"https://openalex.org/keywords/overhead","display_name":"Overhead (engineering)","score":0.4163999855518341}],"concepts":[{"id":"https://openalex.org/C41008148","wikidata":"https://www.wikidata.org/wiki/Q21198","display_name":"Computer science","level":0,"score":0.7907999753952026},{"id":"https://openalex.org/C28855332","wikidata":"https://www.wikidata.org/wiki/Q198099","display_name":"Quantization (signal processing)","level":2,"score":0.5940999984741211},{"id":"https://openalex.org/C173608175","wikidata":"https://www.wikidata.org/wiki/Q232661","display_name":"Parallel computing","level":1,"score":0.5928000211715698},{"id":"https://openalex.org/C43521106","wikidata":"https://www.wikidata.org/wiki/Q2165493","display_name":"Pipeline (software)","level":2,"score":0.5824000239372253},{"id":"https://openalex.org/C11413529","wikidata":"https://www.wikidata.org/wiki/Q8366","display_name":"Algorithm","level":1,"score":0.5406000018119812},{"id":"https://openalex.org/C26517878","wikidata":"https://www.wikidata.org/wiki/Q228039","display_name":"Key (lock)","level":2,"score":0.5012000203132629},{"id":"https://openalex.org/C167927819","wikidata":"https://www.wikidata.org/wiki/Q1930567","display_name":"Shuffling","level":2,"score":0.4966000020503998},{"id":"https://openalex.org/C77618280","wikidata":"https://www.wikidata.org/wiki/Q1155772","display_name":"Scheme (mathematics)","level":2,"score":0.43970000743865967},{"id":"https://openalex.org/C2776214188","wikidata":"https://www.wikidata.org/wiki/Q408386","display_name":"Inference","level":2,"score":0.4359999895095825},{"id":"https://openalex.org/C2778119891","wikidata":"https://www.wikidata.org/wiki/Q477690","display_name":"CUDA","level":2,"score":0.4255000054836273},{"id":"https://openalex.org/C2779960059","wikidata":"https://www.wikidata.org/wiki/Q7113681","display_name":"Overhead (engineering)","level":2,"score":0.4163999855518341},{"id":"https://openalex.org/C80444323","wikidata":"https://www.wikidata.org/wiki/Q2878974","display_name":"Theoretical computer science","level":1,"score":0.36629998683929443},{"id":"https://openalex.org/C111335779","wikidata":"https://www.wikidata.org/wiki/Q3454686","display_name":"Reduction (mathematics)","level":2,"score":0.35089999437332153},{"id":"https://openalex.org/C113775141","wikidata":"https://www.wikidata.org/wiki/Q428691","display_name":"Computer engineering","level":1,"score":0.3474000096321106},{"id":"https://openalex.org/C82876162","wikidata":"https://www.wikidata.org/wiki/Q17096504","display_name":"Latency (audio)","level":2,"score":0.3456000089645386},{"id":"https://openalex.org/C45374587","wikidata":"https://www.wikidata.org/wiki/Q12525525","display_name":"Computation","level":2,"score":0.33629998564720154},{"id":"https://openalex.org/C459310","wikidata":"https://www.wikidata.org/wiki/Q117801","display_name":"Computational science","level":1,"score":0.3319000005722046},{"id":"https://openalex.org/C94375191","wikidata":"https://www.wikidata.org/wiki/Q11205","display_name":"Arithmetic","level":1,"score":0.290800005197525},{"id":"https://openalex.org/C50630238","wikidata":"https://www.wikidata.org/wiki/Q971505","display_name":"General-purpose computing on graphics processing units","level":3,"score":0.27129998803138733},{"id":"https://openalex.org/C206729178","wikidata":"https://www.wikidata.org/wiki/Q2271896","display_name":"Scheduling (production processes)","level":2,"score":0.26499998569488525},{"id":"https://openalex.org/C2989134064","wikidata":"https://www.wikidata.org/wiki/Q288510","display_name":"Execution time","level":2,"score":0.2574999928474426},{"id":"https://openalex.org/C178489894","wikidata":"https://www.wikidata.org/wiki/Q8789","display_name":"Cryptography","level":2,"score":0.2542000114917755},{"id":"https://openalex.org/C57273362","wikidata":"https://www.wikidata.org/wiki/Q576722","display_name":"Decoding methods","level":2,"score":0.2533000111579895},{"id":"https://openalex.org/C46637626","wikidata":"https://www.wikidata.org/wiki/Q6693015","display_name":"Low latency (capital markets)","level":2,"score":0.25200000405311584},{"id":"https://openalex.org/C155281189","wikidata":"https://www.wikidata.org/wiki/Q3518150","display_name":"Tensor (intrinsic definition)","level":2,"score":0.2502000033855438}],"mesh":[],"locations_count":1,"locations":[{"id":"doi:10.1145/3774934.3786423","is_oa":false,"landing_page_url":"https://doi.org/10.1145/3774934.3786423","pdf_url":null,"source":null,"license":null,"license_id":null,"version":"publishedVersion","is_accepted":true,"is_published":true,"raw_source_name":"Proceedings of the 31st ACM SIGPLAN Annual Symposium on Principles and Practice of Parallel Programming","raw_type":"proceedings-article"}],"best_oa_location":null,"sustainable_development_goals":[],"awards":[],"funders":[],"has_content":{"grobid_xml":false,"pdf":false},"content_urls":null,"referenced_works_count":6,"referenced_works":["https://openalex.org/W2963122961","https://openalex.org/W2965862774","https://openalex.org/W2979826702","https://openalex.org/W4321466385","https://openalex.org/W4408029577","https://openalex.org/W6893007900"],"related_works":[],"abstract_inverted_index":{"While":[0],"Large":[1],"Language":[2],"Models":[3],"(LLMs)":[4],"are":[5],"widely":[6],"adopted,":[7],"their":[8],"massive":[9],"parameter":[10],"size":[11],"constrains":[12],"practical":[13],"deployment.":[14],"A":[15],"common":[16],"solution":[17],"is":[18],"clustering-based":[19],"non-uniform":[20],"quantization,":[21],"which":[22],"effectively":[23,100],"compresses":[24],"models":[25],"to":[26,55,71,128],"as":[27,29],"low":[28],"3":[30],"bits":[31],"per":[32],"weight":[33],"while":[34],"preserving":[35],"high":[36],"accuracy.":[37],"However,":[38],"instead":[39],"of":[40,118,126],"accelerating":[41],"memory-bound":[42],"LLM":[43],"inference,":[44,113],"the":[45,63],"memory":[46,73],"reduction":[47],"paradoxically":[48],"often":[49],"causes":[50],"a":[51,68,84,95],"significant":[52],"slowdown":[53],"due":[54],"dequantization":[56],"overhead":[57],"and":[58,93,106,123],"GPU":[59],"underutilization.":[60],"To":[61],"address":[62],"issue,":[64],"we":[65],"propose":[66],"Quantix,":[67],"framework":[69],"designed":[70],"convert":[72],"savings":[74],"into":[75],"inference":[76],"speedups.":[77],"Quantix":[78,109],"applies":[79],"two":[80],"key":[81],"optimizations:":[82],"(1)":[83],"hardware-aligned":[85],"bit":[86],"shuffling":[87],"scheme":[88],"for":[89],"efficient":[90],"data":[91],"access,":[92],"(2)":[94],"fused":[96],"dequantization-multiplication":[97],"pipeline":[98],"that":[99],"maps":[101],"workloads":[102],"on":[103,134],"both":[104],"CUDA":[105],"Tensor":[107],"Cores.":[108],"enables":[110],"high-throughput":[111],"batched":[112],"delivering":[114],"average":[115],"kernel-level":[116],"speedups":[117,125],"4.82\u00d7":[119],"over":[120,130],"FP16":[121],"cuBLAS":[122],"end-to-end":[124],"up":[127],"11.46\u00d7":[129],"state-of-the-art":[131],"quantization":[132],"methods":[133],"NVIDIA":[135],"L40":[136],"GPUs.":[137]},"counts_by_year":[],"updated_date":"2026-04-09T08:11:56.329763","created_date":"2026-01-29T00:00:00"}