RelatedSessioninGTC2022MerlinHugeCTR:由GPU加速的推荐系统训练和推理[S41352]-MinseokLee|NVIDIAMERLIN推荐系统团队高级经理MerlinHugeCTR:使用GPUEmbedding缓存的分布式分层推理参数服务器[S41126]-YingcanWei,FanYu,MatthiasLanger|NVIDIABuildingandDeployingRecommenderSystemsQuicklyandEasilywithNVIDIAMerlin[S41119]–EvenOldridge,SeniorManager,MerlinRecommenderSystemsTeam,NVIDIAGettingstartedHugeCTR@NVIDIA.com:https://developer.nvidia.com/nvidia-merlin/hugectrHugeCTR@GitHub:https://github.com/NVIDIA-Merlin/HugeCTRSuccessstories•LeadingDesignandDevelopmentoftheAdvertisingRecommenderSystematTencent:AnInterviewwithXiangtingKong•Meituan/OptimizingMeituan’sMachineLearningPlatform:AnInterviewwithJunHuangLearnmoreaboutHugeCTR•AcceleratingEmbeddingwiththeHugeCTRTensorFlowEmbeddingPlugin•HugeCTRSeriesPart1:ScalingandAcceleratinglargeDeepLearningRecommenderSystems(CN)HugeCTR系列第1部分：扩展和加速大型深度学习推荐系统•HugeCTRSeriesPart2:TraininglargeDeepLearningRecommenderModelswithMerlinHugeCTR’sPythonAPIs(CN)HugeCTR系列第2部分：使用MerlinHugeCTR的PythonAPI训练大型深度学习推荐模型•HugeCTRParameterServerSeriesPart1:IntroductiontoHierarchicalParameterServerWeareHiring(FullTime&Intern):C++Engineer,CUDAEngineer,RecommendationSystemAlgorithmResearcherPleaseemailyourResumeto:sh-recruitment@nvidia.comHugeCTRResourcesMERLINHUGECTR:GPU-ACCELERATEDRECOMMENDERSYSTEMTRAININGANDINFERENCEJERRYSHI3SOCIALMEDIADIGITALADVERTISINGE-COMMERCEDIGITALCONTENTRECOMMENDERSTHEPERSONALIZATIONENGINEOFTHEINTERNET4.3BActiveUsers4.3BWatchVideosOnline3.7BShopOnline4.7BInternetUsers“Already,35percentofwhatconsumerspurchaseonAmazonand75percentofwhattheywatchonNetflixcomefromproductrecommendationsbasedonsuchalgorithms.”Source:McKinsey4NVTabularPipelinesareslowandcomplexChallengeSolutionInferenceTrainingDataLoadingETLUsingcommonitem-by-itemloadingcanbeslowHighthroughputtorankmoreitemsisdifficultwhilemaintaininglowlatencyEmbeddingtablesoflargedeeplearningrecommendersystemscanexceedmemoryGPU-acceleratedandeasy-to-u...
