{"id":"https://openalex.org/W7225165335","doi":"https://doi.org/10.5281/zenodo.23924376","title":"The Byte Tax: Decode Speed of Byte-Level Language Models on a Consumer GPU","display_name":"The Byte Tax: Decode Speed of Byte-Level Language Models on a Consumer GPU","publication_year":2026,"publication_date":"2026-11-03","ids":{"openalex":"https://openalex.org/W7225165335","doi":"https://doi.org/10.5281/zenodo.23924376"},"language":"en","primary_location":{"id":"doi:10.5281/zenodo.23924376","doi":"https://doi.org/10.5281/zenodo.23924376","is_oa":true,"landing_page_url":"https://doi.org/10.5281/zenodo.23924376","pdf_url":null,"source":{"id":"https://openalex.org/S4306400562","display_name":"Zenodo (CERN European Organization for Nuclear Research)","issn_l":null,"issn":null,"is_oa":true,"is_in_doaj":false,"is_core":false,"listed_in":[],"host_organization":"https://openalex.org/I67311998","host_organization_name":"European Organization for Nuclear Research","host_organization_lineage":["https://openalex.org/I67311998"],"host_organization_lineage_names":[],"type":"repository"},"license":"cc-by","license_id":"https://openalex.org/licenses/cc-by","version":null,"is_accepted":false,"is_published":false,"raw_source_name":null,"raw_type":"Preprint"},"volume":null,"type":"preprint","indexed_in":["datacite"],"open_access":{"is_oa":true,"oa_status":"green","oa_url":"https://doi.org/10.5281/zenodo.23924376","any_repository_has_fulltext":true},"authorships":[{"author_position":"first","author":{"id":"https://openalex.org/A5126487546","display_name":"Fabio-Eric Rempel","orcid":null,"observed_orcids":[]},"institutions":[],"countries":[],"is_corresponding":true,"raw_author_name":"Rempel, Fabio-Eric","raw_affiliation_strings":["Independent researcher"],"raw_orcid":"https://orcid.org/0009-0008-8447-6159","affiliations":[{"raw_affiliation_string":"Independent researcher","institution_ids":[]}]}],"institutions":[],"countries_distinct_count":0,"institutions_distinct_count":0,"corresponding_author_ids":["https://openalex.org/A5126487546"],"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":null,"last_page":null},"is_retracted":false,"is_paratext":false,"is_xpac":false,"primary_topic":{"id":"https://openalex.org/T10054","display_name":"Parallel Computing and Optimization Techniques","score":0.6161999702453613,"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":0.6161999702453613,"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":0.21580000221729279,"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/T10028","display_name":"Topic Modeling","score":0.06040000170469284,"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/decoding-speed","display_name":"decoding speed","score":0.7129999995231628},{"id":"https://openalex.org/keywords/olmo","display_name":"OLMo","score":0.6499999761581421},{"id":"https://openalex.org/keywords/consumer-gpus","display_name":"consumer GPUs","score":0.44999998807907104},{"id":"https://openalex.org/keywords/cuda-graphs","display_name":"CUDA Graphs","score":0.6869999766349792},{"id":"https://openalex.org/keywords/weight-quantization","display_name":"weight quantization","score":0.7979999780654907},{"id":"https://openalex.org/keywords/speculative-decoding","display_name":"speculative decoding","score":0.7360000014305115},{"id":"https://openalex.org/keywords/memory-bandwidth","display_name":"memory bandwidth","score":0.593999981880188},{"id":"https://openalex.org/keywords/kernel-profiling","display_name":"kernel profiling","score":0.43700000643730164},{"id":"https://openalex.org/keywords/int8-quantization","display_name":"INT8 quantization","score":0.5429999828338623},{"id":"https://openalex.org/keywords/bfloat16","display_name":"bfloat16","score":0.6890000104904175},{"id":"https://openalex.org/keywords/patch-boundaries","display_name":"patch boundaries","score":0.31200000643730164}],"concepts":[{"id":"https://openalex.org/C43364308","wikidata":"https://www.wikidata.org/wiki/Q8799","display_name":"Byte","level":2,"score":0.8367999792098999},{"id":"https://openalex.org/C41008148","wikidata":"https://www.wikidata.org/wiki/Q21198","display_name":"Computer science","level":0,"score":0.79830002784729},{"id":"https://openalex.org/C48145219","wikidata":"https://www.wikidata.org/wiki/Q1335365","display_name":"Security token","level":2,"score":0.6913999915122986},{"id":"https://openalex.org/C57273362","wikidata":"https://www.wikidata.org/wiki/Q576722","display_name":"Decoding methods","level":2,"score":0.5135999917984009},{"id":"https://openalex.org/C173608175","wikidata":"https://www.wikidata.org/wiki/Q232661","display_name":"Parallel computing","level":1,"score":0.48750001192092896},{"id":"https://openalex.org/C137293760","wikidata":"https://www.wikidata.org/wiki/Q3621696","display_name":"Language model","level":2,"score":0.4765999913215637},{"id":"https://openalex.org/C2779960059","wikidata":"https://www.wikidata.org/wiki/Q7113681","display_name":"Overhead (engineering)","level":2,"score":0.4616999924182892},{"id":"https://openalex.org/C74193536","wikidata":"https://www.wikidata.org/wiki/Q574844","display_name":"Kernel (algebra)","level":2,"score":0.4381999969482422},{"id":"https://openalex.org/C2778119891","wikidata":"https://www.wikidata.org/wiki/Q477690","display_name":"CUDA","level":2,"score":0.42500001192092896},{"id":"https://openalex.org/C118505674","wikidata":"https://www.wikidata.org/wiki/Q42586063","display_name":"Encoder","level":2,"score":0.4016000032424927},{"id":"https://openalex.org/C11413529","wikidata":"https://www.wikidata.org/wiki/Q8366","display_name":"Algorithm","level":1,"score":0.3553999960422516},{"id":"https://openalex.org/C82876162","wikidata":"https://www.wikidata.org/wiki/Q17096504","display_name":"Latency (audio)","level":2,"score":0.35339999198913574},{"id":"https://openalex.org/C28855332","wikidata":"https://www.wikidata.org/wiki/Q198099","display_name":"Quantization (signal processing)","level":2,"score":0.3262999951839447},{"id":"https://openalex.org/C111919701","wikidata":"https://www.wikidata.org/wiki/Q9135","display_name":"Operating system","level":1,"score":0.30300000309944153},{"id":"https://openalex.org/C164155591","wikidata":"https://www.wikidata.org/wiki/Q2067766","display_name":"Satisfiability modulo theories","level":2,"score":0.27219998836517334},{"id":"https://openalex.org/C80444323","wikidata":"https://www.wikidata.org/wiki/Q2878974","display_name":"Theoretical computer science","level":1,"score":0.26100000739097595}],"mesh":[],"locations_count":2,"locations":[{"id":"doi:10.5281/zenodo.23924376","doi":"https://doi.org/10.5281/zenodo.23924376","is_oa":true,"landing_page_url":"https://doi.org/10.5281/zenodo.23924376","pdf_url":null,"source":{"id":"https://openalex.org/S4306400562","display_name":"Zenodo (CERN European Organization for Nuclear Research)","issn_l":null,"issn":null,"is_oa":true,"is_in_doaj":false,"is_core":false,"listed_in":[],"host_organization":"https://openalex.org/I67311998","host_organization_name":"European Organization for Nuclear Research","host_organization_lineage":["https://openalex.org/I67311998"],"host_organization_lineage_names":[],"type":"repository"},"license":"cc-by","license_id":"https://openalex.org/licenses/cc-by","version":null,"is_accepted":false,"is_published":false,"raw_source_name":null,"raw_type":"Preprint"},{"id":"doi:10.5281/zenodo.23924377","doi":"https://doi.org/10.5281/zenodo.23924377","is_oa":true,"landing_page_url":"https://doi.org/10.5281/zenodo.23924377","pdf_url":null,"source":{"id":"https://openalex.org/S4306400562","display_name":"Zenodo (CERN European Organization for Nuclear Research)","issn_l":null,"issn":null,"is_oa":true,"is_in_doaj":false,"is_core":false,"listed_in":[],"host_organization":"https://openalex.org/I67311998","host_organization_name":"European Organization for Nuclear Research","host_organization_lineage":["https://openalex.org/I67311998"],"host_organization_lineage_names":[],"type":"repository"},"license":"cc-by","license_id":"https://openalex.org/licenses/cc-by","version":null,"is_accepted":false,"is_published":false,"raw_source_name":null,"raw_type":"Preprint"}],"best_oa_location":{"id":"doi:10.5281/zenodo.23924376","doi":"https://doi.org/10.5281/zenodo.23924376","is_oa":true,"landing_page_url":"https://doi.org/10.5281/zenodo.23924376","pdf_url":null,"source":{"id":"https://openalex.org/S4306400562","display_name":"Zenodo (CERN European Organization for Nuclear Research)","issn_l":null,"issn":null,"is_oa":true,"is_in_doaj":false,"is_core":false,"listed_in":[],"host_organization":"https://openalex.org/I67311998","host_organization_name":"European Organization for Nuclear Research","host_organization_lineage":["https://openalex.org/I67311998"],"host_organization_lineage_names":[],"type":"repository"},"license":"cc-by","license_id":"https://openalex.org/licenses/cc-by","version":null,"is_accepted":false,"is_published":false,"raw_source_name":null,"raw_type":"Preprint"},"sustainable_development_goals":[],"study_designs":[],"awards":[],"funders":[],"has_content":{"grobid_xml":false,"pdf":false},"content_urls":null,"referenced_works_count":1,"referenced_works":["https://openalex.org/W7220840186"],"related_works":["https://openalex.org/W7160835164","https://openalex.org/W7127320815","https://openalex.org/W7220729864","https://openalex.org/W4405433019","https://openalex.org/W7212663828","https://openalex.org/W7225137372","https://openalex.org/W7211874740","https://openalex.org/W7167900054","https://openalex.org/W7224828044","https://openalex.org/W7220840186"],"abstract_inverted_index":{"Byte-level":[0],"language":[1],"models":[2],"remove":[3],"the":[4,19,70,110,115,120,132,139,189,195,203,209,217,230,259,269,274,277,285,292,298,317,320,332,361,367],"tokenizer,":[5],"but":[6],"at":[7,162,207,212,300],"batch":[8,399],"size":[9,400],"1":[10],"they":[11,22],"decode":[12,89],"more":[13,127],"slowly":[14],"per":[15,66,129],"generated":[16],"byte":[17,28,155,163],"than":[18,131,178,254],"token":[20,144,204],"model":[21,111,145],"come":[23],"from.":[24],"We":[25],"measure":[26],"this":[27,266],"tax":[29,140,191,210],"for":[30,151,220],"Bolmo":[31,62,92,175,250],"1B,":[32,37],"a":[33,42,49,87,98,104,143,154,240],"byteified":[34],"OLMo":[35],"2":[36],"against":[38],"its":[39,78,179,330],"source":[40],"on":[41,310],"consumer":[43],"GPU":[44],"(RTX":[45],"3060":[46,394],"Ti),":[47],"through":[48],"chain":[50],"of":[51,69,100,199,258,280,297,316,335,345,375],"67":[52,260,286],"hypotheses":[53,261,288],"written":[54,328],"down":[55],"before":[56,329],"each":[57,336],"measurement.":[58],"In":[59],"eager":[60],"decoding":[61],"is":[63,72,119,233],"2.36\u00d7":[64],"slower":[65],"byte.":[67],"Most":[68],"gap":[71,99],"launch":[73],"and":[74,82,86,96,112,153,160,187,206,226,238,256,284,319,340,360,371,385],"host":[75],"overhead":[76,232],"in":[77,182,185,265],"per-byte":[79,116],"local":[80,117],"encoder":[81],"decoder:":[83],"CUDA":[84],"graphs":[85],"single-sync":[88],"loop":[90,181],"make":[91],"3.8\u00d7":[93],"faster,":[94],"byte-identically,":[95],"leave":[97],"1.5\u00d7":[101],"that":[102,114,221,229,239],"follows":[103],"memory-bandwidth":[105],"model.":[106],"Weight":[107],"quantization":[108],"confirms":[109],"shows":[113],"stack":[118],"cheapest":[121],"part":[122],"to":[123,141,168,192,247,366],"quantize,":[124],"saving":[125],"4.1\u00d7":[126],"time":[128],"gigabyte":[130],"global":[133],"transformer.":[134],"Speculative":[135],"drafting":[136,157,166,362],"first":[137],"widens":[138],"2.37\u00d7:":[142],"verifies":[146],"several":[147],"drafted":[148,190],"tokens":[149],"almost":[150],"free,":[152],"model\u2019s":[156],"rounds":[158,167],"end":[159],"restart":[161],"granularity.":[164],"Aligning":[165],"patch":[169],"boundaries":[170],"reverses":[171],"this.":[172],"It":[173],"makes":[174],"2.06\u00d7":[176],"faster":[177,253],"plain":[180],"bf16,":[183],"exact":[184],"fp32,":[186],"narrows":[188],"1.36\u00d7,":[193],"below":[194],"undrafted":[196],"1.49\u00d7.":[197],"Optimizations":[198],"weight":[200],"reads":[201],"favour":[202],"model,":[205],"int8":[208],"settles":[211],"1.60\u00d7,":[213],"about":[214],"13%":[215],"above":[216],"bandwidth":[218],"floor":[219],"configuration.":[222],"A":[223],"kernel":[224,234],"census":[225],"microbenchmarks":[227],"show":[228],"remaining":[231],"count,":[235],"not":[236],"synchronization,":[237],"persistent":[241],"megakernel":[242],"would":[243],"gain":[244],"only":[245],"4":[246],"8%.":[248],"Overall":[249],"becomes":[251],"10.3\u00d7":[252],"eager,":[255],"26":[257],"were":[262],"refuted.":[263],"Files":[264],"record":[267,279],"byte_tax_paper.pdf:":[268],"paper":[270,318],"(10":[271],"pages).":[272],"byte_tax_technical_report.pdf:":[273],"technical":[275],"report,":[276],"full":[278],"all":[281,376],"13":[282],"studies":[283],"pre-registered":[287],"with":[289,324],"verdicts,":[290],"including":[291],"negative":[293],"results.":[294],"byteprof-1.0.0.zip:":[295],"snapshot":[296],"code":[299,380],"v1.0.0":[301],"(profiling":[302],"harness,":[303],"kernels,":[304],"configurations,":[305],"scripts,":[306],"tests;":[307],"Apache-2.0).":[308],"Also":[309],"GitHub:":[311],"github.com/fabs133/byteprof.":[312],"latex_sources.zip:":[313],"LaTeX":[314],"sources":[315],"report.":[321],"study_records.zip:":[322],"FINDINGS.md":[323],"every":[325,346],"hypothesis":[326],"as":[327],"measurement,":[331],"task":[333],"specification":[334],"study,":[337],"run":[338,347],"logs":[339],"notes.":[341],"results_core.zip:":[342],"raw":[343],"results":[344],"(reports,":[348],"JSON":[349],"summaries,":[350],"quality":[351],"evaluations).":[352],"results_traces.zip:":[353],"large":[354],"profiler":[355],"traces":[356],"(torch.profiler,":[357],"Nsight":[358],"Systems)":[359],"corpus.":[363],"README_ZENODO.md:":[364],"guide":[365],"files,":[368],"hardware,":[369],"licenses":[370],"reproduction.":[372],"SHA256SUMS.txt:":[373],"checksums":[374],"files":[377],"above.":[378],"Licenses:":[379],"under":[381,387],"Apache-2.0;":[382],"paper,":[383],"report":[384],"data":[386],"CC":[388],"BY":[389],"4.0.":[390],"All":[391],"measurements:":[392],"RTX":[393],"Ti":[395],"(8":[396],"GB),":[397],"Windows,":[398],"1.":[401]},"counts_by_year":[],"updated_date":"2026-11-05T12:25:35.769452","created_date":"2026-11-04T00:00:00"}