{"id":"https://openalex.org/W7106100991","doi":"https://doi.org/10.1109/iccad66269.2025.11240807","title":"MoE-OPU: An FPGA Overlay Processor Leveraging Expert Parallelism for MoE-based Large Language Models","display_name":"MoE-OPU: An FPGA Overlay Processor Leveraging Expert Parallelism for MoE-based Large Language Models","publication_year":2025,"publication_date":"2025-10-26","ids":{"openalex":"https://openalex.org/W7106100991","doi":"https://doi.org/10.1109/iccad66269.2025.11240807"},"language":null,"primary_location":{"id":"doi:10.1109/iccad66269.2025.11240807","is_oa":false,"landing_page_url":"https://doi.org/10.1109/iccad66269.2025.11240807","pdf_url":null,"source":null,"license":null,"license_id":null,"version":"publishedVersion","is_accepted":true,"is_published":true,"raw_source_name":"2025 IEEE/ACM International Conference On Computer Aided Design (ICCAD)","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":null,"display_name":"Shaoqiang Lu","orcid":null},"institutions":[{"id":"https://openalex.org/I183067930","display_name":"Shanghai Jiao Tong University","ror":"https://ror.org/0220qvk04","country_code":"CN","type":"education","lineage":["https://openalex.org/I183067930"]}],"countries":["CN"],"is_corresponding":false,"raw_author_name":"Shaoqiang Lu","raw_affiliation_strings":["Shanghai Jiao Tong University,Shanghai,China"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"Shanghai Jiao Tong University,Shanghai,China","institution_ids":["https://openalex.org/I183067930"]}]},{"author_position":"middle","author":{"id":null,"display_name":"Yangbo Wei","orcid":null},"institutions":[{"id":"https://openalex.org/I183067930","display_name":"Shanghai Jiao Tong University","ror":"https://ror.org/0220qvk04","country_code":"CN","type":"education","lineage":["https://openalex.org/I183067930"]}],"countries":["CN"],"is_corresponding":false,"raw_author_name":"Yangbo Wei","raw_affiliation_strings":["Shanghai Jiao Tong University,Shanghai,China"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"Shanghai Jiao Tong University,Shanghai,China","institution_ids":["https://openalex.org/I183067930"]}]},{"author_position":"middle","author":{"id":null,"display_name":"Junhong Qian","orcid":null},"institutions":[{"id":"https://openalex.org/I76569877","display_name":"Southeast University","ror":"https://ror.org/04ct4d772","country_code":"CN","type":"education","lineage":["https://openalex.org/I76569877"]}],"countries":["CN"],"is_corresponding":false,"raw_author_name":"Junhong Qian","raw_affiliation_strings":["Southeast University,Nanjing,China"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"Southeast University,Nanjing,China","institution_ids":["https://openalex.org/I76569877"]}]},{"author_position":"middle","author":{"id":null,"display_name":"Chen Wu","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Chen Wu","raw_affiliation_strings":["Eastern Institute of Technology,Ningbo,China"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"Eastern Institute of Technology,Ningbo,China","institution_ids":[]}]},{"author_position":"middle","author":{"id":null,"display_name":"Xiao Shi","orcid":null},"institutions":[{"id":"https://openalex.org/I76569877","display_name":"Southeast University","ror":"https://ror.org/04ct4d772","country_code":"CN","type":"education","lineage":["https://openalex.org/I76569877"]}],"countries":["CN"],"is_corresponding":false,"raw_author_name":"Xiao Shi","raw_affiliation_strings":["Southeast University,Nanjing,China"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"Southeast University,Nanjing,China","institution_ids":["https://openalex.org/I76569877"]}]},{"author_position":"last","author":{"id":null,"display_name":"Lei He","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Lei He","raw_affiliation_strings":["Eastern Institute of Technology,Ningbo,China"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"Eastern Institute of Technology,Ningbo,China","institution_ids":[]}]}],"institutions":[],"countries_distinct_count":1,"institutions_distinct_count":6,"corresponding_author_ids":[],"corresponding_institution_ids":[],"apc_list":null,"apc_paid":null,"fwci":0.0,"has_fulltext":false,"cited_by_count":0,"citation_normalized_percentile":{"value":0.64376335,"is_in_top_1_percent":false,"is_in_top_10_percent":false},"cited_by_percentile_year":null,"biblio":{"volume":null,"issue":null,"first_page":"1","last_page":"9"},"is_retracted":false,"is_paratext":false,"is_xpac":false,"primary_topic":{"id":"https://openalex.org/T14347","display_name":"Big Data and Digital Economy","score":0.18080000579357147,"subfield":{"id":"https://openalex.org/subfields/1710","display_name":"Information Systems"},"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/T14347","display_name":"Big Data and Digital Economy","score":0.18080000579357147,"subfield":{"id":"https://openalex.org/subfields/1710","display_name":"Information Systems"},"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":0.10570000112056732,"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/T10036","display_name":"Advanced Neural Network Applications","score":0.08569999784231186,"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"}}],"keywords":[{"id":"https://openalex.org/keywords/scalability","display_name":"Scalability","score":0.5781000256538391},{"id":"https://openalex.org/keywords/field-programmable-gate-array","display_name":"Field-programmable gate array","score":0.5514000058174133},{"id":"https://openalex.org/keywords/scheduling","display_name":"Scheduling (production processes)","score":0.4542999863624573},{"id":"https://openalex.org/keywords/language-model","display_name":"Language model","score":0.4219000041484833},{"id":"https://openalex.org/keywords/edge-device","display_name":"Edge device","score":0.4050000011920929},{"id":"https://openalex.org/keywords/quantization","display_name":"Quantization (signal processing)","score":0.40380001068115234},{"id":"https://openalex.org/keywords/security-token","display_name":"Security token","score":0.40139999985694885},{"id":"https://openalex.org/keywords/software-deployment","display_name":"Software deployment","score":0.39640000462532043},{"id":"https://openalex.org/keywords/inference","display_name":"Inference","score":0.3695000112056732}],"concepts":[{"id":"https://openalex.org/C41008148","wikidata":"https://www.wikidata.org/wiki/Q21198","display_name":"Computer science","level":0,"score":0.843999981880188},{"id":"https://openalex.org/C48044578","wikidata":"https://www.wikidata.org/wiki/Q727490","display_name":"Scalability","level":2,"score":0.5781000256538391},{"id":"https://openalex.org/C42935608","wikidata":"https://www.wikidata.org/wiki/Q190411","display_name":"Field-programmable gate array","level":2,"score":0.5514000058174133},{"id":"https://openalex.org/C206729178","wikidata":"https://www.wikidata.org/wiki/Q2271896","display_name":"Scheduling (production processes)","level":2,"score":0.4542999863624573},{"id":"https://openalex.org/C137293760","wikidata":"https://www.wikidata.org/wiki/Q3621696","display_name":"Language model","level":2,"score":0.4219000041484833},{"id":"https://openalex.org/C120314980","wikidata":"https://www.wikidata.org/wiki/Q180634","display_name":"Distributed computing","level":1,"score":0.40639999508857727},{"id":"https://openalex.org/C173608175","wikidata":"https://www.wikidata.org/wiki/Q232661","display_name":"Parallel computing","level":1,"score":0.4052000045776367},{"id":"https://openalex.org/C138236772","wikidata":"https://www.wikidata.org/wiki/Q25098575","display_name":"Edge device","level":3,"score":0.4050000011920929},{"id":"https://openalex.org/C28855332","wikidata":"https://www.wikidata.org/wiki/Q198099","display_name":"Quantization (signal processing)","level":2,"score":0.40380001068115234},{"id":"https://openalex.org/C48145219","wikidata":"https://www.wikidata.org/wiki/Q1335365","display_name":"Security token","level":2,"score":0.40139999985694885},{"id":"https://openalex.org/C105339364","wikidata":"https://www.wikidata.org/wiki/Q2297740","display_name":"Software deployment","level":2,"score":0.39640000462532043},{"id":"https://openalex.org/C2776214188","wikidata":"https://www.wikidata.org/wiki/Q408386","display_name":"Inference","level":2,"score":0.3695000112056732},{"id":"https://openalex.org/C26517878","wikidata":"https://www.wikidata.org/wiki/Q228039","display_name":"Key (lock)","level":2,"score":0.35199999809265137},{"id":"https://openalex.org/C118524514","wikidata":"https://www.wikidata.org/wiki/Q173212","display_name":"Computer architecture","level":1,"score":0.3319000005722046},{"id":"https://openalex.org/C2776834041","wikidata":"https://www.wikidata.org/wiki/Q25346349","display_name":"Execution model","level":2,"score":0.3303999900817871},{"id":"https://openalex.org/C108010975","wikidata":"https://www.wikidata.org/wiki/Q500094","display_name":"Pruning","level":2,"score":0.3077000081539154},{"id":"https://openalex.org/C50644808","wikidata":"https://www.wikidata.org/wiki/Q192776","display_name":"Artificial neural network","level":2,"score":0.3068000078201294},{"id":"https://openalex.org/C2780149590","wikidata":"https://www.wikidata.org/wiki/Q7302742","display_name":"Reconfigurability","level":2,"score":0.29490000009536743},{"id":"https://openalex.org/C149635348","wikidata":"https://www.wikidata.org/wiki/Q193040","display_name":"Embedded system","level":1,"score":0.29409998655319214},{"id":"https://openalex.org/C41661131","wikidata":"https://www.wikidata.org/wiki/Q220764","display_name":"Interrupt","level":3,"score":0.28200000524520874},{"id":"https://openalex.org/C2776257435","wikidata":"https://www.wikidata.org/wiki/Q1576430","display_name":"Bandwidth (computing)","level":2,"score":0.27799999713897705},{"id":"https://openalex.org/C20136886","wikidata":"https://www.wikidata.org/wiki/Q749647","display_name":"Interoperability","level":2,"score":0.2728999853134155},{"id":"https://openalex.org/C34165917","wikidata":"https://www.wikidata.org/wiki/Q188267","display_name":"Programming paradigm","level":2,"score":0.26600000262260437},{"id":"https://openalex.org/C29202148","wikidata":"https://www.wikidata.org/wiki/Q287260","display_name":"Resource allocation","level":2,"score":0.26080000400543213},{"id":"https://openalex.org/C31258907","wikidata":"https://www.wikidata.org/wiki/Q1301371","display_name":"Computer network","level":1,"score":0.26080000400543213},{"id":"https://openalex.org/C125411270","wikidata":"https://www.wikidata.org/wiki/Q18653","display_name":"Encoding (memory)","level":2,"score":0.2606000006198883},{"id":"https://openalex.org/C43521106","wikidata":"https://www.wikidata.org/wiki/Q2165493","display_name":"Pipeline (software)","level":2,"score":0.25679999589920044},{"id":"https://openalex.org/C101468663","wikidata":"https://www.wikidata.org/wiki/Q1620158","display_name":"Modular design","level":2,"score":0.2556999921798706},{"id":"https://openalex.org/C113775141","wikidata":"https://www.wikidata.org/wiki/Q428691","display_name":"Computer engineering","level":1,"score":0.2502000033855438}],"mesh":[],"locations_count":1,"locations":[{"id":"doi:10.1109/iccad66269.2025.11240807","is_oa":false,"landing_page_url":"https://doi.org/10.1109/iccad66269.2025.11240807","pdf_url":null,"source":null,"license":null,"license_id":null,"version":"publishedVersion","is_accepted":true,"is_published":true,"raw_source_name":"2025 IEEE/ACM International Conference On Computer Aided Design (ICCAD)","raw_type":"proceedings-article"}],"best_oa_location":null,"sustainable_development_goals":[],"awards":[],"funders":[],"has_content":{"pdf":false,"grobid_xml":false},"content_urls":null,"referenced_works_count":27,"referenced_works":["https://openalex.org/W2979455536","https://openalex.org/W3008408165","https://openalex.org/W3034100311","https://openalex.org/W3206749822","https://openalex.org/W3212505503","https://openalex.org/W4211085526","https://openalex.org/W4308083513","https://openalex.org/W4386396242","https://openalex.org/W4386902741","https://openalex.org/W4389162698","https://openalex.org/W4389162736","https://openalex.org/W4391125526","https://openalex.org/W4392746077","https://openalex.org/W4393407064","https://openalex.org/W4393578753","https://openalex.org/W4393949386","https://openalex.org/W4401211642","https://openalex.org/W4401414720","https://openalex.org/W4402667057","https://openalex.org/W4402671950","https://openalex.org/W4404133735","https://openalex.org/W4404134154","https://openalex.org/W4404955001","https://openalex.org/W4407953477","https://openalex.org/W4407953517","https://openalex.org/W4409285540","https://openalex.org/W4415799085"],"related_works":[],"abstract_inverted_index":{"The":[0],"advent":[1],"of":[2,33,59,159,169,237,243],"Large":[3],"Language":[4],"Models":[5],"(LLMs)":[6],"like":[7],"DeepSeek,":[8],"empowered":[9],"by":[10,56,112,120,144],"the":[11,57,104,117,167,201,208,235],"Mixture-of-Experts":[12],"(MoE)":[13],"architecture,":[14],"has":[15],"driven":[16],"significant":[17],"advancements":[18,233],"across":[19],"diverse":[20],"applications.":[21],"However,":[22],"a":[23,30,135,162,173,196,213],"critical":[24],"challenge":[25],"arises":[26],"during":[27],"inference:":[28],"Only":[29],"small":[31],"fraction":[32],"experts":[34,143],"are":[35],"activated,":[36],"causing":[37],"severe":[38],"token":[39,216],"allocation":[40],"imbalances":[41],"among":[42],"experts.":[43],"This":[44],"inefficiency":[45],"poses":[46],"substantial":[47],"storage":[48],"and":[49,68,107,149,180,190,222,228],"computational":[50],"burdens":[51],"on":[52,195,246],"resource-constrained":[53],"devices,":[54],"exacerbated":[55],"lack":[58],"optimization":[60],"strategies":[61],"that":[62,88,176,205],"integrate":[63],"expert":[64,113],"usage-aware":[65],"parameter":[66,118],"pruning":[67],"parallel":[69,90,185],"scheduling,":[70],"ultimately":[71],"leading":[72],"to":[73,122,187,219],"suboptimal":[74],"resource":[75],"utilization.":[76],"To":[77],"address":[78],"these":[79],"limitations,":[80],"we":[81,98],"propose":[82],"MoE-OPU,":[83],"an":[84,154],"FPGA-based":[85],"overlay":[86],"processor":[87],"optimizes":[89],"MoE":[91],"inference":[92],"through":[93],"three":[94],"key":[95],"innovations.":[96],"First,":[97],"introduce":[99],"N:M":[100],"sparsity":[101],"(1:4/2:4/4:8/6:8/8:8)":[102],"in":[103],"MLP":[105],"layers":[106],"mixed-precision":[108,181],"quantization":[109],"(BF16/FP8/INT4)":[110],"guided":[111],"activation":[114,147],"frequency,":[115],"reducing":[116],"size":[119],"up":[121],"2.76\u00d7":[123],"while":[124],"maintaining":[125],"model":[126,203],"accuracy":[127],"(only":[128],"1.53%":[129],"average":[130,155],"drop":[131],"after":[132],"fine-tuning).":[133],"Second,":[134],"lightweight":[136],"prediction":[137,156],"network":[138],"dynamically":[139],"predicts":[140],"next-layer":[141],"\"hot\"":[142],"analyzing":[145],"historical":[146],"patterns":[148],"current":[150],"hidden":[151],"states,":[152],"achieving":[153],"hit":[157],"rate":[158],"83.4%.":[160],"Third,":[161],"reconfigurable":[163],"multi-core":[164],"architecture":[165],"maximizes":[166],"utilization":[168],"HBM":[170],"bandwidth":[171],"via":[172],"systolic":[174],"array":[175],"natively":[177],"supports":[178],"sparse":[179],"computations,":[182],"coupled":[183],"with":[184,200],"concatenation":[186],"balance":[188],"compute":[189],"memory":[191],"efficiency.":[192],"Experimental":[193],"results":[194],"Xilinx":[197],"V80":[198],"FPGA":[199],"DeepSeek-V2-lite":[202],"demonstrate":[204],"MoE-OPU":[206,225],"outperforms":[207],"NVIDIA":[209],"A100":[210],"GPU,":[211],"delivering":[212],"6.78\u00d7":[214],"higher":[215],"throughput.":[217],"Compared":[218],"RTX":[220],"4090":[221],"U200":[223],"FPGA,":[224],"achieves":[226],"13.37\u00d7":[227],"7.85\u00d7":[229],"improvements,":[230],"respectively.":[231],"These":[232],"highlight":[234],"potential":[236],"algorithm-hardware":[238],"co-design":[239],"for":[240],"scalable":[241],"deployment":[242],"MoE-based":[244],"LLMs":[245],"edge":[247],"devices.":[248]},"counts_by_year":[],"updated_date":"2026-06-11T09:08:48.828518","created_date":"2025-11-20T00:00:00"}