NVIDIA 助力构建安全和高效的推荐系统.pdf

1、RECOMMENDERS - THE ENGINE OF THE INTERNETBilions of Users- Trilions of Items100Sof millions of etail items Amazon 8 Alibabarecommenders10003 of movies Netflix recommenderslapuauuoal saunll B AyHods sSuos Jo suoyu Jo S.0L10s of millions of books - Amazon recommenderBillions of Tik Toka YT videos- TT&

2、 YTrecommenderBillions of websites Google search rankSo much news!Il s Google 8 FB news recommendersShMDL#page#DEEP LEARNING RECOMMENDERS ARE A DIRECT PATHTO INCREASED REVENUESGather DataModel ComplexityTraining FrequencyClick-Through-RateRevenueLARGER DATASETSDEEP LEARNING MODELSFRESHERMODELSBETTER

3、 RECOMMENDATIONSIMPROVEOUTPERFORMINCREASESIMPROVESRECOMMENDATION QUALITYTRADITIONAL METHODSPROBABILITY OF PURCHASEENGAGEMENTAND RETENTIONnWIDIA#page#INTRODUCING DGX A100The Universal AI System- Data Analytics， Training and Inference18o0NVSwcNVME SSD#page#NEW MULTI-INSTANCE GPU (MIG）Optimize GPU Util

4、ization Expand Access to More Users with Guaranteed Quality of ServiceUp To 7 GPU Instances In a Single A100: DedicatedAmberSM，Memory，L2 cache，Bandwidth for hardwareQoS8isolationSimultaneous Workload Execution With GuaranteedQuality Of Service: AIL MIG instances run inPUGPUSTCCPUGPJparallel with pre

5、dictable throughput 8 latencyGPUMomGPU MomGPU MSmGPU MamGFUMmGPUMamGPUMomRight Sized GPU Allocation:Different sized MIGinstances based on targetworkloadsFlexibility to run any type ofworkload on a MGinstanceDiverse Deployment Environments: Supported withBare metal，Docker Kubernetes，Virtualized Env.S

6、ee MIG in action Running Al Inference and Mixed WorkloadsShMbr#page#NEED FOR INFERENCING AT THE EDGEREAL-TIMEAIAT THE EDGECORE DATACENTERBILLIONSOFSENSORSTHE NVIDIAEGXPLATFORMRETRAININGAND REPORTINGFromJe13INSICHTSANOMALYALER#page#MODERN APPS DEMAND ADVANCED NETWORKINGMonolithicArchitectureMicroserv