GPU 在向量搜索中的應用.pdf

1、GPU在向量搜索中的應用金海ZillizZILLI2820207#page#Speaker bio金海ZILLI2R&D DirectorProject LeaderCareer historyEducation2020 Milvus.All righhts reserved#page#Unlock the treasure of unstructured dataAlalgorithms transform image，video，voice，naturallanguageinto vectors，and enable understandingandutilization ofunstru

2、ctured data at scale.mUnstructured dataDeep learning modelsEmbedding vectorsKnowledge，insight，$#page#The flow-based Al applicationsVideoThe most popularwayVoiceExtractExtractFlexibleframcsmodeltagsEasytocomposeweb-based UISample pipelinesImageThe challengeData fragmentationVsual modelVGG.gEmbeddings

3、EmbeddingsAttributesVisualVoice2020 Milvus.All righhts reserved#page#The unstructured data service (UDS) for AlUnstructured Dataimage，video，voice，naturallanguageSearchInsertModel Inference RuntimeInference LayerStoreTensorRT.ONNX RT.TFRT.ctc.InsertSearchMilvusObject URIObjectVectorsAttributesData Se

4、rvice LayerStorageHigh dense +SparsexperimentalVector IDMultimodalScoringon roadmapon roadmapOutputResult Setimage，video，voice，natural language#page#Milvus:The big pictureQueryProcessing EngineBuffer PoolSchedulerANNSCollaborative QuenIndexM-FAISS,Mi-Annoytag/structureddataSDK/ Web APIResultFiestop-

5、K resultReducerScoringMultimodalapp specificSegmentSegmentMetadataquery objSelectioninsert objX86:supporisSSE4.2，AVX2,AVX512GPU:Pareorlaler.CUDA10.0imieroarehlX86ArmrlateNewIndexArm:requirosaarch64IndexFilesFileGPUOtherVarious ProcessorsStorage Tier#page#Vectors are differentNumbersVectorsArithmetic

6、 operationSimilarity(cg.Euclidcan distance）-+d（4B）=a-a）OperatiSimilaritycomparisonNumber comparisonTopK（A）=argmin（d（AB）91-106-101-5123891045672020 Milvus.All rights reserve#page#Faiss-GPUTraining data0.32,0.21.0.510.52,0.11.0.60IVF Flatf.0.91,0.33.0.13CentroidsClustersGPUVector data0.32,0.21.0.510.5

7、2,0.11.0.60有0.91,0.33.0.132020Milvus.Allrighsreserved#page#Faiss - Multiple GPUs supportFaiss-IndexShardsFaiss- IndexReplicasGPUOGPU1GPUOGPU1#page#Other LimitationsThroughputSM-0SM-1SM-（N-1）HighTopK = 2048L2 CacheGlobal MemoryLOWGPUPCIE2048Thread queueWarp queueInsertDDW。W2TooTo.Tom-1Merging Network

8、Lane0W.WLane1T:T；CheckLane31WT31oT31To1 m-1WWaD.RegisterRegisterGlobal Memory2020 Miivus.All rights reserved#page#Challenges - CPU-GPU co-processingFaiss-IVFQueryIndex0.32,0.21,0.23.0.51）Step1CentroidsQuery0.32,0.21,0.23，.0.51Step2Clusters2020 Milvus.All righhts reserved#page#Challenges - CPU-GPU co

9、-processingFaiss-IVF Indexnlist2x-5x speed up!nlist* dimension*4GPUCentroidsbytesCPUClusters#page#Performance ComparisonCPUvS CPU+GPUOS:Centos Linux release 7.6.1810（Core）CPU:Intel(R）Xeon(R）CPUE5-2678v32.50GGPU:Tesla P100*2Memory:503GBDataset:sift1B（http:/corpus-texmex.irisa.fr）metric_type:L2nlist:1

10、6384nprobe:32200吉NuberO QueriesOCPU+GPLOCPU2020 Milvus.All righhts reserved#page#Our Journey2018.102019.042019.06MivsThe most activeAl projects in1st5eedLinux foundationMis.1The deaUse8Joine年A2020.102019.102020.032020 Milvus.All righhts reserved#page#Community ConferenceMilvusmmunityonf2020#page#Cur

11、rent Process45136098703122CommitsGitHub starsForksContributors16190K70+400+UsersReleasesDockerHub downloadsMeetupsandevents2020 Milvus.All rights reserve#page#lmage search for companytrademarkImages55millionimagesCompany TrademarkSearchelapsedtime商標查詢20ms on cloud GPUserverEncoderVGG (fine tuned）Obj

12、ectURIObjectSearchMilvusStorageVector IDTrademarklmageCompany Info2020 Milvus.All righhts reserved#page#Intelligent writing assistantCorpus Datanatural languageWriting IntentonData CleansingEncoderTextCNNResultExAnauto-generatedEncoderessayInferSentObjectURObjectMilvusStorageVectorID#page#Pharmaceut

13、ical molecule analysisMolecular Formula800millionmoleculesCC(=O)Nclccc(S(=O）（=ONCC(=ON2CCS(=0)CC2)cc1Searchelapsedtime:Encoder500 ms on single serverRDKitMolecular fingerprint1024bits00001100.10000000MilvusTanimotosimilarityMolecularSubstructureCandidate ListsimilaritySuperstructure2020 Milvus.All r

14、ighhts reserved#page#Users400+users in 10months貝殼P再來能t99mWMPS七中國電信X服Q系M科技6微博等節額動平安科技魔快手0IY愛奇藝A然物不接母蘇寧易購M二湖起科技4399BEsTmnnSBESNEA治明Ultrapouer08捷信圖微拍堂HaG中用研學院新華智春沿科技店中千O匯D銀加水箱0！Galiboo焦點科技2望石智慧Microsofttokopedic金VC5GtwwFleneKUANG-CHI光啟拍云SlprivatBankST88.851日東學甲WISE GBITS6ssyasteroxylonURANMINNOGiENiden

15、tvNCECuniviewJSCCEIECe.5om深度行G恒科技流喜回奇安信企意意稀定科技阿里云Flat面將石動SENSOROFINNOS荔校平安人壽X3TRANSSIONG京東A7DHgccAEEBOSS直聘途虎養車#page#Zilliz:WhoweareOpen-source software company basedin ShanghaiSoftwareewelopmenMain contributor of Milvus projectScene風山MachineMotndLeaming2020 Milvus.All righhts reserved#page#We are hi

16、ringFind our positions in ChinaC+backenddeveloperAl algorithm engineerFrontend developerProduct managerProject managerFind our positions in USCloud infrastructure engineer/developerOpen-source evangelist（US）Developeradvocate（US）Community manager(US）Anda lot more.You may also contact 2020Mivus.Allrighsreserved#page#Thanks口 Speaker Contact口WeChat:13917405881Mail: