{"id":"https://openalex.org/W7204129100","doi":"https://doi.org/10.48550/arxiv.2608.21240","title":"SPICE: Speculative Prefetching with Low-Rank Expert Surrogates and Heterogeneous Orchestration for MoE Inference Acceleration","display_name":"SPICE: Speculative Prefetching with Low-Rank Expert Surrogates and Heterogeneous Orchestration for MoE Inference Acceleration","publication_year":2026,"publication_date":"2026-08-21","ids":{"openalex":"https://openalex.org/W7204129100","doi":"https://doi.org/10.48550/arxiv.2608.21240"},"language":"en","primary_location":{"id":"pmh:oai:arXiv.org:2608.21240","doi":null,"is_oa":true,"landing_page_url":"https://arxiv.org/abs/2608.21240","pdf_url":"https://arxiv.org/pdf/2608.21240","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":null,"license_id":null,"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/2608.21240","any_repository_has_fulltext":true},"authorships":[{"author_position":"first","author":{"id":"https://openalex.org/A5019985626","display_name":"Yongxiang Lyu","orcid":null,"observed_orcids":[]},"institutions":[{"id":"https://openalex.org/I137902535","display_name":"North Carolina State University","ror":"https://ror.org/04tj63d06","country_code":"US","type":"education","lineage":["https://openalex.org/I137902535"]}],"countries":["US"],"is_corresponding":false,"raw_author_name":"Lyu, Yongxiang","raw_affiliation_strings":["North Carolina State University United States"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"North Carolina State University United States","institution_ids":["https://openalex.org/I137902535"]}]},{"author_position":"middle","author":{"id":"https://openalex.org/A5147946715","display_name":"Ning Li","orcid":null,"observed_orcids":[]},"institutions":[{"id":"https://openalex.org/I162868743","display_name":"Tianjin University","ror":"https://ror.org/012tb2g32","country_code":"CN","type":"education","lineage":["https://openalex.org/I162868743"]}],"countries":["CN"],"is_corresponding":false,"raw_author_name":"Li, Ning","raw_affiliation_strings":["Tianjin University China"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"Tianjin University China","institution_ids":["https://openalex.org/I162868743"]}]},{"author_position":"last","author":{"id":"https://openalex.org/A5102696683","display_name":"Bonian Jia","orcid":null,"observed_orcids":[]},"institutions":[{"id":"https://openalex.org/I57206974","display_name":"New York University","ror":"https://ror.org/0190ak572","country_code":"US","type":"education","lineage":["https://openalex.org/I57206974"]}],"countries":["US"],"is_corresponding":false,"raw_author_name":"Jia, Bonian","raw_affiliation_strings":["New York University United States"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"New York University United States","institution_ids":["https://openalex.org/I57206974"]}]}],"institutions":[],"countries_distinct_count":2,"institutions_distinct_count":3,"corresponding_author_ids":[],"corresponding_institution_ids":[],"apc_list":null,"apc_paid":null,"fwci":null,"has_fulltext":true,"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":0.9916999936103821,"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":0.9916999936103821,"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":0.0031999999191612005,"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":0.0005000000237487257,"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/mixture-of-experts","display_name":"mixture of experts","score":0.9210000038146973},{"id":"https://openalex.org/keywords/speculative-prefetching","display_name":"speculative prefetching","score":0.9139999747276306},{"id":"https://openalex.org/keywords/large-language-model-inference","display_name":"large language model inference","score":0.6200000047683716},{"id":"https://openalex.org/keywords/gpu-memory-constraints","display_name":"GPU memory constraints","score":0.4880000054836273}],"concepts":[{"id":"https://openalex.org/C41008148","wikidata":"https://www.wikidata.org/wiki/Q21198","display_name":"Computer science","level":0,"score":0.8532000184059143},{"id":"https://openalex.org/C2776214188","wikidata":"https://www.wikidata.org/wiki/Q408386","display_name":"Inference","level":2,"score":0.6823999881744385},{"id":"https://openalex.org/C68339613","wikidata":"https://www.wikidata.org/wiki/Q1549489","display_name":"Speedup","level":2,"score":0.6394000053405762},{"id":"https://openalex.org/C46743427","wikidata":"https://www.wikidata.org/wiki/Q1341685","display_name":"Inference engine","level":3,"score":0.5450999736785889},{"id":"https://openalex.org/C155512373","wikidata":"https://www.wikidata.org/wiki/Q287450","display_name":"Residual","level":2,"score":0.5127999782562256},{"id":"https://openalex.org/C133588205","wikidata":"https://www.wikidata.org/wiki/Q28455645","display_name":"Instruction prefetch","level":3,"score":0.48730000853538513},{"id":"https://openalex.org/C105002631","wikidata":"https://www.wikidata.org/wiki/Q4833645","display_name":"Subject-matter expert","level":3,"score":0.44920000433921814},{"id":"https://openalex.org/C7366592","wikidata":"https://www.wikidata.org/wiki/Q1255620","display_name":"Dram","level":2,"score":0.4438999891281128},{"id":"https://openalex.org/C82876162","wikidata":"https://www.wikidata.org/wiki/Q17096504","display_name":"Latency (audio)","level":2,"score":0.41130000352859497},{"id":"https://openalex.org/C199168358","wikidata":"https://www.wikidata.org/wiki/Q3367000","display_name":"Orchestration","level":3,"score":0.3781000077724457},{"id":"https://openalex.org/C74912251","wikidata":"https://www.wikidata.org/wiki/Q6815727","display_name":"Memory footprint","level":2,"score":0.3650999963283539},{"id":"https://openalex.org/C129916263","wikidata":"https://www.wikidata.org/wiki/Q1141183","display_name":"Backward chaining","level":4,"score":0.35120001435279846},{"id":"https://openalex.org/C190839683","wikidata":"https://www.wikidata.org/wiki/Q2448197","display_name":"Train","level":2,"score":0.3334999978542328},{"id":"https://openalex.org/C119857082","wikidata":"https://www.wikidata.org/wiki/Q2539","display_name":"Machine learning","level":1,"score":0.33149999380111694},{"id":"https://openalex.org/C117896860","wikidata":"https://www.wikidata.org/wiki/Q11376","display_name":"Acceleration","level":2,"score":0.3280999958515167},{"id":"https://openalex.org/C64270927","wikidata":"https://www.wikidata.org/wiki/Q206924","display_name":"PCI Express","level":3,"score":0.3188999891281128},{"id":"https://openalex.org/C52723943","wikidata":"https://www.wikidata.org/wiki/Q1127410","display_name":"Serialization","level":2,"score":0.3125},{"id":"https://openalex.org/C173608175","wikidata":"https://www.wikidata.org/wiki/Q232661","display_name":"Parallel computing","level":1,"score":0.30079999566078186},{"id":"https://openalex.org/C172430144","wikidata":"https://www.wikidata.org/wiki/Q17111997","display_name":"Symmetric multiprocessor system","level":2,"score":0.29750001430511475},{"id":"https://openalex.org/C48145219","wikidata":"https://www.wikidata.org/wiki/Q1335365","display_name":"Security token","level":2,"score":0.29600000381469727},{"id":"https://openalex.org/C154945302","wikidata":"https://www.wikidata.org/wiki/Q11660","display_name":"Artificial intelligence","level":1,"score":0.29600000381469727},{"id":"https://openalex.org/C187191949","wikidata":"https://www.wikidata.org/wiki/Q1138496","display_name":"Profiling (computer programming)","level":2,"score":0.2928999960422516},{"id":"https://openalex.org/C46681722","wikidata":"https://www.wikidata.org/wiki/Q4725589","display_name":"Alias","level":2,"score":0.2922999858856201},{"id":"https://openalex.org/C2777538425","wikidata":"https://www.wikidata.org/wiki/Q1154300","display_name":"Cannibalization","level":2,"score":0.29100000858306885},{"id":"https://openalex.org/C113775141","wikidata":"https://www.wikidata.org/wiki/Q428691","display_name":"Computer engineering","level":1,"score":0.28679999709129333},{"id":"https://openalex.org/C168065819","wikidata":"https://www.wikidata.org/wiki/Q845566","display_name":"Debugging","level":2,"score":0.28459998965263367},{"id":"https://openalex.org/C2777472644","wikidata":"https://www.wikidata.org/wiki/Q16968992","display_name":"Approximate inference","level":3,"score":0.28450000286102295},{"id":"https://openalex.org/C207685749","wikidata":"https://www.wikidata.org/wiki/Q2088941","display_name":"Domain knowledge","level":2,"score":0.2809000015258789},{"id":"https://openalex.org/C169590947","wikidata":"https://www.wikidata.org/wiki/Q47506","display_name":"Compiler","level":2,"score":0.2800999879837036},{"id":"https://openalex.org/C58328972","wikidata":"https://www.wikidata.org/wiki/Q184609","display_name":"Expert system","level":2,"score":0.2793999910354614},{"id":"https://openalex.org/C28719098","wikidata":"https://www.wikidata.org/wiki/Q44946","display_name":"Point (geometry)","level":2,"score":0.2732999920845032},{"id":"https://openalex.org/C3826847","wikidata":"https://www.wikidata.org/wiki/Q188768","display_name":"FLOPS","level":2,"score":0.27090001106262207},{"id":"https://openalex.org/C79403827","wikidata":"https://www.wikidata.org/wiki/Q3988","display_name":"Real-time computing","level":1,"score":0.27070000767707825},{"id":"https://openalex.org/C3913047","wikidata":"https://www.wikidata.org/wiki/Q1956265","display_name":"sync","level":3,"score":0.2678999900817871},{"id":"https://openalex.org/C160234255","wikidata":"https://www.wikidata.org/wiki/Q812535","display_name":"Bayesian inference","level":3,"score":0.26350000500679016},{"id":"https://openalex.org/C172191483","wikidata":"https://www.wikidata.org/wiki/Q1071806","display_name":"Provisioning","level":2,"score":0.26109999418258667},{"id":"https://openalex.org/C18762648","wikidata":"https://www.wikidata.org/wiki/Q42213","display_name":"Work (physics)","level":2,"score":0.25999999046325684},{"id":"https://openalex.org/C36503486","wikidata":"https://www.wikidata.org/wiki/Q11235244","display_name":"Domain (mathematical analysis)","level":2,"score":0.25870001316070557},{"id":"https://openalex.org/C134652429","wikidata":"https://www.wikidata.org/wiki/Q1052698","display_name":"Jitter","level":2,"score":0.25529998540878296}],"mesh":[],"locations_count":2,"locations":[{"id":"pmh:oai:arXiv.org:2608.21240","doi":null,"is_oa":true,"landing_page_url":"https://arxiv.org/abs/2608.21240","pdf_url":"https://arxiv.org/pdf/2608.21240","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":null,"license_id":null,"version":"submittedVersion","is_accepted":false,"is_published":false,"raw_source_name":null,"raw_type":"text"},{"id":"doi:10.48550/arxiv.2608.21240","doi":"https://doi.org/10.48550/arxiv.2608.21240","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2608.21240","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":null,"license_id":null,"version":null,"is_accepted":false,"is_published":false,"raw_source_name":null,"raw_type":"Preprint"}],"best_oa_location":{"id":"pmh:oai:arXiv.org:2608.21240","doi":null,"is_oa":true,"landing_page_url":"https://arxiv.org/abs/2608.21240","pdf_url":"https://arxiv.org/pdf/2608.21240","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":null,"license_id":null,"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/W7204129100.pdf","grobid_xml":"https://content.openalex.org/works/W7204129100.grobid-xml"},"referenced_works_count":2,"referenced_works":["https://openalex.org/W4401660192","https://openalex.org/W7127959502"],"related_works":["https://openalex.org/W7090930907","https://openalex.org/W7140083998","https://openalex.org/W7171708010","https://openalex.org/W7222956968","https://openalex.org/W7106172866","https://openalex.org/W7164891892","https://openalex.org/W7161029323","https://openalex.org/W4415105204","https://openalex.org/W7135051689","https://openalex.org/W4417065185"],"abstract_inverted_index":{"Mixture-of-Experts":[0],"(MoE)":[1],"models":[2],"are":[3,107],"increasingly":[4],"used":[5],"in":[6,132,153],"LLMs":[7],"because":[8],"sparse":[9],"activation":[10],"decouples":[11],"model":[12,72],"capacity":[13],"from":[14],"compute":[15],"cost.":[16],"However,":[17],"the":[18,33,75,90,110,127],"large":[19],"expert":[20,38,58,113,117],"parameter":[21],"footprint":[22],"often":[23],"exceeds":[24],"GPU":[25,136,145],"memory":[26],"capacity,":[27],"making":[28],"inference":[29],"latency":[30],"dominated":[31],"by":[32,109],"host-to-device":[34],"PCIe":[35],"transfers":[36],"for":[37,52],"loading.":[39],"To":[40],"address":[41],"these":[42],"challenges,":[43],"this":[44],"paper":[45],"presents":[46],"SPICE,":[47],"a":[48,69,80,100],"speculative":[49,94],"prefetching":[50],"framework":[51],"MoE":[53,77,166],"offloading":[54,167],"that":[55,164],"combines":[56],"lightweight":[57,70],"prediction":[59],"with":[60,74,114,134,159],"confidence-aware":[61,81],"CPU-GPU":[62,102],"orchestration.":[63],"On":[64,89],"one":[65],"hand,":[66,92],"SPICE":[67,97,147],"builds":[68],"draft":[71],"aligned":[73],"target":[76],"architecture,":[78],"using":[79],"adaptive":[82],"lookahead":[83],"algorithm":[84],"to":[85,99,126,150],"prefetch":[86],"high-confidence":[87],"experts.":[88],"other":[91],"when":[93],"predictions":[95],"miss,":[96],"switches":[98],"cost-aware":[101],"heterogeneous":[103],"orchestration:":[104],"low-confidence":[105],"misses":[106,178],"approximated":[108],"resident":[111],"shared":[112],"low":[115],"rank":[116],"(LoRE)":[118],"surrogates,":[119],"while":[120],"exact":[121,183,187],"residual":[122,188],"work":[123,189],"is":[124],"offloaded":[125],"CPU":[128],"and":[129,141,185],"executed":[130],"asynchronously":[131],"parallel":[133],"ongoing":[135],"computation.":[137],"Evaluated":[138],"on":[139],"DeepSeek-V2-Lite":[140],"Qwen2-57B-A14B":[142],"across":[143],"diverse":[144],"platforms,":[146],"achieves":[148],"up":[149],"3.12":[151],"speedup":[152],"Time":[154],"Per":[155],"Output":[156],"Token":[157],"(TPOT)":[158],"minimal":[160],"quality":[161],"loss,":[162],"showing":[163],"effective":[165],"requires":[168],"not":[169],"only":[170],"predicting":[171],"future":[172],"experts,":[173],"but":[174],"also":[175],"deciding":[176],"which":[177,181],"deserve":[179],"approximation,":[180],"require":[182],"recovery,":[184],"where":[186],"should":[190],"execute.":[191]},"counts_by_year":[],"updated_date":"2026-11-07T11:25:07.637180","created_date":"2026-08-25T00:00:00"}