{"id":"https://openalex.org/W7172503556","doi":"https://doi.org/10.48550/arxiv.2608.03741","title":"When Does Disaggregation Pay? Simulating Prefill--Decode--Attention--FFN Specialization for Agentic LLM Inference","display_name":"When Does Disaggregation Pay? Simulating Prefill--Decode--Attention--FFN Specialization for Agentic LLM Inference","publication_year":2026,"publication_date":"2026-08-04","ids":{"openalex":"https://openalex.org/W7172503556","doi":"https://doi.org/10.48550/arxiv.2608.03741"},"language":"en","primary_location":{"id":"pmh:oai:arXiv.org:2608.03741","doi":null,"is_oa":true,"landing_page_url":"https://arxiv.org/abs/2608.03741","pdf_url":"https://arxiv.org/pdf/2608.03741","source":{"id":"https://openalex.org/S4306400194","display_name":"arXiv (Cornell University)","issn_l":"2331-8422","issn":["2331-8422"],"is_oa":true,"is_in_doaj":false,"is_core":false,"listed_in":[],"host_organization":"https://openalex.org/I205783295","host_organization_name":"Cornell University","host_organization_lineage":["https://openalex.org/I205783295"],"host_organization_lineage_names":[],"type":"repository"},"license":"cc-by","license_id":"https://openalex.org/licenses/cc-by","version":"submittedVersion","is_accepted":false,"is_published":false,"raw_source_name":null,"raw_type":"text"},"volume":null,"type":"preprint","is_traditional":true,"indexed_in":["arxiv","datacite"],"open_access":{"is_oa":true,"oa_status":"green","oa_url":"https://arxiv.org/pdf/2608.03741","any_repository_has_fulltext":true},"authorships":[{"author_position":"first","author":{"id":"https://openalex.org/A5134027946","display_name":"Przemyslaw Forys","orcid":null,"observed_orcids":[]},"institutions":[{"id":"https://openalex.org/I47508984","display_name":"Imperial College London","ror":"https://ror.org/041kmwe10","country_code":"GB","type":"education","lineage":["https://openalex.org/I47508984"]}],"countries":["GB"],"is_corresponding":false,"raw_author_name":"Forys, Przemyslaw","raw_affiliation_strings":["Imperial College London , UK"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"Imperial College London , UK","institution_ids":["https://openalex.org/I47508984"]}]},{"author_position":"middle","author":{"id":"https://openalex.org/A5134085235","display_name":"Haoran Wu","orcid":null,"observed_orcids":[]},"institutions":[{"id":"https://openalex.org/I241749","display_name":"University of Cambridge","ror":"https://ror.org/013meh722","country_code":"GB","type":"education","lineage":["https://openalex.org/I241749"]}],"countries":["GB"],"is_corresponding":false,"raw_author_name":"Wu, Haoran","raw_affiliation_strings":["University of Cambridge , UK"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"University of Cambridge , UK","institution_ids":["https://openalex.org/I241749"]}]},{"author_position":"middle","author":{"id":"https://openalex.org/A5124239666","display_name":"Can Xiao","orcid":null,"observed_orcids":[]},"institutions":[{"id":"https://openalex.org/I47508984","display_name":"Imperial College London","ror":"https://ror.org/041kmwe10","country_code":"GB","type":"education","lineage":["https://openalex.org/I47508984"]}],"countries":["GB"],"is_corresponding":false,"raw_author_name":"Xiao, Can","raw_affiliation_strings":["Imperial College London , UK"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"Imperial College London , UK","institution_ids":["https://openalex.org/I47508984"]}]},{"author_position":"middle","author":{"id":"https://openalex.org/A5124271622","display_name":"Jiayi Nie","orcid":null,"observed_orcids":[]},"institutions":[{"id":"https://openalex.org/I241749","display_name":"University of Cambridge","ror":"https://ror.org/013meh722","country_code":"GB","type":"education","lineage":["https://openalex.org/I241749"]}],"countries":["GB"],"is_corresponding":false,"raw_author_name":"Nie, Jiayi","raw_affiliation_strings":["University of Cambridge , UK"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"University of Cambridge , UK","institution_ids":["https://openalex.org/I241749"]}]},{"author_position":"middle","author":{"id":"https://openalex.org/A5144388821","display_name":"Tony Liu","orcid":null,"observed_orcids":[]},"institutions":[{"id":"https://openalex.org/I47508984","display_name":"Imperial College London","ror":"https://ror.org/041kmwe10","country_code":"GB","type":"education","lineage":["https://openalex.org/I47508984"]}],"countries":["GB"],"is_corresponding":false,"raw_author_name":"Liu, Tony","raw_affiliation_strings":["Imperial College London , UK"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"Imperial College London , UK","institution_ids":["https://openalex.org/I47508984"]}]},{"author_position":"middle","author":{"id":"https://openalex.org/A5009099496","display_name":"Rika Antonova","orcid":"https://orcid.org/0000-0002-3018-2445","observed_orcids":["https://orcid.org/0000-0002-3018-2445"]},"institutions":[{"id":"https://openalex.org/I241749","display_name":"University of Cambridge","ror":"https://ror.org/013meh722","country_code":"GB","type":"education","lineage":["https://openalex.org/I241749"]}],"countries":["GB"],"is_corresponding":false,"raw_author_name":"Antonova, Rika","raw_affiliation_strings":["University of Cambridge , UK"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"University of Cambridge , UK","institution_ids":["https://openalex.org/I241749"]}]},{"author_position":"middle","author":{"id":"https://openalex.org/A5138275051","display_name":"Timothy Jones","orcid":"https://orcid.org/0000-0001-7372-6345","observed_orcids":["https://orcid.org/0000-0001-7372-6345"]},"institutions":[{"id":"https://openalex.org/I241749","display_name":"University of Cambridge","ror":"https://ror.org/013meh722","country_code":"GB","type":"education","lineage":["https://openalex.org/I241749"]}],"countries":["GB"],"is_corresponding":false,"raw_author_name":"Jones, Timothy","raw_affiliation_strings":["University of Cambridge , UK"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"University of Cambridge , UK","institution_ids":["https://openalex.org/I241749"]}]},{"author_position":"middle","author":{"id":"https://openalex.org/A5128724690","display_name":"Robert Mullins","orcid":null,"observed_orcids":[]},"institutions":[{"id":"https://openalex.org/I241749","display_name":"University of Cambridge","ror":"https://ror.org/013meh722","country_code":"GB","type":"education","lineage":["https://openalex.org/I241749"]}],"countries":["GB"],"is_corresponding":false,"raw_author_name":"Mullins, Robert","raw_affiliation_strings":["University of Cambridge , UK"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"University of Cambridge , UK","institution_ids":["https://openalex.org/I241749"]}]},{"author_position":"middle","author":{"id":"https://openalex.org/A5057940557","display_name":"Wayne W. Luk","orcid":"https://orcid.org/0000-0002-6750-927X","observed_orcids":["https://orcid.org/0000-0002-6750-927X"]},"institutions":[{"id":"https://openalex.org/I47508984","display_name":"Imperial College London","ror":"https://ror.org/041kmwe10","country_code":"GB","type":"education","lineage":["https://openalex.org/I47508984"]}],"countries":["GB"],"is_corresponding":false,"raw_author_name":"Luk, Wayne","raw_affiliation_strings":["Imperial College London , UK"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"Imperial College London , UK","institution_ids":["https://openalex.org/I47508984"]}]},{"author_position":"middle","author":{"id":"https://openalex.org/A5113840125","display_name":"Aaron Zhao","orcid":null,"observed_orcids":[]},"institutions":[{"id":"https://openalex.org/I47508984","display_name":"Imperial College London","ror":"https://ror.org/041kmwe10","country_code":"GB","type":"education","lineage":["https://openalex.org/I47508984"]}],"countries":["GB"],"is_corresponding":false,"raw_author_name":"Zhao, Aaron","raw_affiliation_strings":["Imperial College London , UK"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"Imperial College London , UK","institution_ids":["https://openalex.org/I47508984"]}]},{"author_position":"last","author":{"id":"https://openalex.org/A5096154656","display_name":"George A. Constantinides","orcid":"https://orcid.org/0000-0002-0201-310X","observed_orcids":["https://orcid.org/0000-0002-0201-310X"]},"institutions":[{"id":"https://openalex.org/I47508984","display_name":"Imperial College London","ror":"https://ror.org/041kmwe10","country_code":"GB","type":"education","lineage":["https://openalex.org/I47508984"]}],"countries":["GB"],"is_corresponding":false,"raw_author_name":"Constantinides, George A.","raw_affiliation_strings":["Imperial College London , UK"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"Imperial College London , UK","institution_ids":["https://openalex.org/I47508984"]}]}],"institutions":[],"countries_distinct_count":1,"institutions_distinct_count":2,"corresponding_author_ids":[],"corresponding_institution_ids":[],"apc_list":null,"apc_paid":null,"fwci":null,"has_fulltext":true,"cited_by_count":2,"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":0.9560999870300293,"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":0.9560999870300293,"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/T10036","display_name":"Advanced Neural Network Applications","score":0.007799999788403511,"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/T10904","display_name":"Embedded Systems Design Techniques","score":0.006399999838322401,"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/heterogeneous-hardware","display_name":"heterogeneous hardware","score":0.6320000290870667},{"id":"https://openalex.org/keywords/model-quantization","display_name":"model quantization","score":0.640999972820282}],"concepts":[{"id":"https://openalex.org/C2776214188","wikidata":"https://www.wikidata.org/wiki/Q408386","display_name":"Inference","level":2,"score":0.8203999996185303},{"id":"https://openalex.org/C41008148","wikidata":"https://www.wikidata.org/wiki/Q21198","display_name":"Computer science","level":0,"score":0.7195000052452087},{"id":"https://openalex.org/C177264268","wikidata":"https://www.wikidata.org/wiki/Q1514741","display_name":"Set (abstract data type)","level":2,"score":0.5960999727249146},{"id":"https://openalex.org/C157764524","wikidata":"https://www.wikidata.org/wiki/Q1383412","display_name":"Throughput","level":3,"score":0.5408999919891357},{"id":"https://openalex.org/C168167062","wikidata":"https://www.wikidata.org/wiki/Q1117970","display_name":"Component (thermodynamics)","level":2,"score":0.5117999911308289},{"id":"https://openalex.org/C2778476105","wikidata":"https://www.wikidata.org/wiki/Q628539","display_name":"Workload","level":2,"score":0.48010000586509705},{"id":"https://openalex.org/C66882249","wikidata":"https://www.wikidata.org/wiki/Q169336","display_name":"Homogeneous","level":2,"score":0.453000009059906},{"id":"https://openalex.org/C2777472644","wikidata":"https://www.wikidata.org/wiki/Q16968992","display_name":"Approximate inference","level":3,"score":0.37869998812675476},{"id":"https://openalex.org/C123657996","wikidata":"https://www.wikidata.org/wiki/Q12271","display_name":"Architecture","level":2,"score":0.35839998722076416},{"id":"https://openalex.org/C154945302","wikidata":"https://www.wikidata.org/wiki/Q11660","display_name":"Artificial intelligence","level":1,"score":0.31610000133514404},{"id":"https://openalex.org/C165696696","wikidata":"https://www.wikidata.org/wiki/Q11287","display_name":"Exploit","level":2,"score":0.31200000643730164},{"id":"https://openalex.org/C120314980","wikidata":"https://www.wikidata.org/wiki/Q180634","display_name":"Distributed computing","level":1,"score":0.30889999866485596},{"id":"https://openalex.org/C158600405","wikidata":"https://www.wikidata.org/wiki/Q5054566","display_name":"Causal inference","level":2,"score":0.2994999885559082},{"id":"https://openalex.org/C119857082","wikidata":"https://www.wikidata.org/wiki/Q2539","display_name":"Machine learning","level":1,"score":0.2946999967098236},{"id":"https://openalex.org/C67953723","wikidata":"https://www.wikidata.org/wiki/Q192525","display_name":"Workstation","level":2,"score":0.28450000286102295},{"id":"https://openalex.org/C2776207758","wikidata":"https://www.wikidata.org/wiki/Q5303302","display_name":"Downstream (manufacturing)","level":2,"score":0.26409998536109924},{"id":"https://openalex.org/C2776359362","wikidata":"https://www.wikidata.org/wiki/Q2145286","display_name":"Representation (politics)","level":3,"score":0.25459998846054077}],"mesh":[],"locations_count":2,"locations":[{"id":"pmh:oai:arXiv.org:2608.03741","doi":null,"is_oa":true,"landing_page_url":"https://arxiv.org/abs/2608.03741","pdf_url":"https://arxiv.org/pdf/2608.03741","source":{"id":"https://openalex.org/S4306400194","display_name":"arXiv (Cornell University)","issn_l":"2331-8422","issn":["2331-8422"],"is_oa":true,"is_in_doaj":false,"is_core":false,"listed_in":[],"host_organization":"https://openalex.org/I205783295","host_organization_name":"Cornell University","host_organization_lineage":["https://openalex.org/I205783295"],"host_organization_lineage_names":[],"type":"repository"},"license":"cc-by","license_id":"https://openalex.org/licenses/cc-by","version":"submittedVersion","is_accepted":false,"is_published":false,"raw_source_name":null,"raw_type":"text"},{"id":"doi:10.48550/arxiv.2608.03741","doi":"https://doi.org/10.48550/arxiv.2608.03741","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2608.03741","pdf_url":null,"source":{"id":"https://openalex.org/S4306400194","display_name":"arXiv (Cornell University)","issn_l":"2331-8422","issn":["2331-8422"],"is_oa":true,"is_in_doaj":false,"is_core":false,"listed_in":[],"host_organization":"https://openalex.org/I205783295","host_organization_name":"Cornell University","host_organization_lineage":["https://openalex.org/I205783295"],"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":"pmh:oai:arXiv.org:2608.03741","doi":null,"is_oa":true,"landing_page_url":"https://arxiv.org/abs/2608.03741","pdf_url":"https://arxiv.org/pdf/2608.03741","source":{"id":"https://openalex.org/S4306400194","display_name":"arXiv (Cornell University)","issn_l":"2331-8422","issn":["2331-8422"],"is_oa":true,"is_in_doaj":false,"is_core":false,"listed_in":[],"host_organization":"https://openalex.org/I205783295","host_organization_name":"Cornell University","host_organization_lineage":["https://openalex.org/I205783295"],"host_organization_lineage_names":[],"type":"repository"},"license":"cc-by","license_id":"https://openalex.org/licenses/cc-by","version":"submittedVersion","is_accepted":false,"is_published":false,"raw_source_name":null,"raw_type":"text"},"sustainable_development_goals":[],"study_designs":[],"awards":[],"funders":[],"has_content":{"pdf":true,"grobid_xml":true},"content_urls":{"pdf":"https://content.openalex.org/works/W7172503556.pdf","grobid_xml":"https://content.openalex.org/works/W7172503556.grobid-xml"},"referenced_works_count":11,"referenced_works":["https://openalex.org/W4387725429","https://openalex.org/W4393578753","https://openalex.org/W4401211642","https://openalex.org/W4401211704","https://openalex.org/W4405903187","https://openalex.org/W4414939355","https://openalex.org/W7075994024","https://openalex.org/W7081570314","https://openalex.org/W7138268496","https://openalex.org/W7155016287","https://openalex.org/W7214194744"],"related_works":["https://openalex.org/W7162010250","https://openalex.org/W4415254670","https://openalex.org/W7162644759","https://openalex.org/W7166649151","https://openalex.org/W4407569479","https://openalex.org/W4396788676","https://openalex.org/W7131909639","https://openalex.org/W7125195775","https://openalex.org/W4415254273","https://openalex.org/W4413213050"],"abstract_inverted_index":{"Agentic":[0],"inference":[1,6,28],"now":[2,56],"dominates":[3],"the":[4,26,75,85,89,161,192,207],"LLM":[5],"landscape,":[7],"requiring":[8],"LLMs":[9],"to":[10,58,173,176],"actively":[11],"engage":[12],"in":[13,98],"multi-turn":[14],"interactions":[15],"with":[16,69,79,179],"tool-calling":[17],"capabilities.":[18,47],"This":[19],"introduces":[20],"a":[21,49,51,99,109,149,218],"more":[22],"complex":[23],"workload":[24],"for":[25,95,113,153,195],"underlying":[27],"system:":[29],"serving":[30,71,157,169,178],"stages":[31],"such":[32,73],"as":[33,74],"prefill":[34],"and":[35,41,45,81,131,135,182,212],"decode":[36],"exhibit":[37],"substantially":[38],"different":[39,199],"behaviors":[40],"demand":[42],"distinct":[43],"compute":[44],"memory-bandwidth":[46],"As":[48],"result,":[50],"single":[52],"homogeneous":[53],"GPU":[54],"system":[55,101],"struggles":[57],"support":[59],"agentic":[60,156],"inference,":[61],"motivating":[62],"an":[63],"industry":[64],"shift":[65],"toward":[66],"heterogeneous":[67,100,155],"systems":[68],"disaggregated":[70,114,126],"capabilities,":[72],"emerging":[76],"Vera-Rubin":[77],"platform":[78],"GPUs":[80,181],"Groq":[82],"LPUs.":[83],"However,":[84],"question":[86],"of":[87,163,220],"what":[88],"optimal":[90],"hardware":[91],"should":[92],"look":[93],"like":[94],"each":[96],"component":[97],"remains":[102],"underexplored.":[103],"To":[104],"this":[105],"end,":[106],"we":[107,147],"propose":[108],"novel":[110],"simulation":[111,121,151],"framework":[112,152],"serving,":[115],"termed":[116],"\\textbf{HeteroPanacea},":[117],"that":[118],"enables":[119],"system-level":[120],"across":[122,198],"three":[123,145],"dimensions:":[124],"1)":[125],"quantization,":[127],"2)":[128],"automated":[129],"intra-":[130],"inter-device":[132],"parallelization":[133],"scheduling,":[134],"3)":[136],"PDAF":[137],"(prefill-decode-attention-FFN)":[138],"NPU":[139],"architectural":[140],"heterogeneity.":[141],"By":[142],"combining":[143],"these":[144],"axes,":[146],"provide":[148],"cross-stack":[150],"future":[154],"systems.":[158],"We":[159,204],"confirm":[160],"benefit":[162],"Prefill":[164,186],"Decode":[165,187],"disaggregation,":[166],"simulating":[167],"increased":[168],"throughput":[170,197],"by":[171,216],"up":[172],"75\\%":[174],"compared":[175],"traditional":[177],"current":[180],"demonstrate":[183],"4":[184],"way":[185],"Attention":[188],"FFN":[189],"disaggregation":[190,215],"is":[191],"most":[193],"consistent":[194],"increasing":[196],"models,":[200],"assuming":[201],"custom":[202],"NPUs.":[203],"also":[205],"investigate":[206],"relationship":[208],"between":[209],"model":[210],"architecture":[211],"gain":[213],"from":[214],"running":[217],"set":[219],"ablation":[221],"studies.":[222]},"counts_by_year":[{"year":2026,"cited_by_count":2}],"updated_date":"2026-11-07T11:25:07.637180","created_date":"2026-08-06T00:00:00"}