《Bringing AI Everywhere.pdf》由会员分享,可在线阅读,更多相关《Bringing AI Everywhere.pdf(15页珍藏版)》请在三个皮匠报告上搜索。
1、Bringing AI Everywhere Anurag HandaGenerative AIMachine LearningDeep LearningChallengesAI EverywhereThe rapid growth of AIThe AI landscape2Rapidly growing number of methods,capabilities,data types and sizes,and infrastructure requirements to run AI ComplexityIncreasing costs due to increased compute
2、 demand as AI becomes more widely adopted and consumedCostNeed for speed along with many steps and skill sets required to get AI from proof of concepts to production in a scalable,sustainable processOperationalizingActivating sensitive or regulated data globally while remaining secure and compliantD
3、ata security and privacy Ensuring AI technology advances responsibly,ethically and equitably with a comprehensive approach that lowers risks,improves lives and optimizes benefitsHuman impactBringing AI Everywhere 3Accelerate the WorkloadStreamline the AI WorkflowSimplify AI Infrastructure CloudClien
4、t AI SpecificGeneral PurposeAI SoftwareScalable Systems&SolutionsFoundational Silicon&Software Training&Fine-TuningInference&Deployment SmallModelsUnlock the AI Continuum Large ModelsNovel Applications4Comprehensive Portfolio that meets the needs of EveryoneAI ContinuumModel CreationTraining/Fine-Tu
5、ningDeploymentEdge InferenceLocalized Inference(Client)Data PrepCLOUD&ENTERPRISECLIENT&WORKSTATIONEDGEAI ModelsDirectMLHugging FaceML applicationsAI Accelerator Roadmap Intel Gaudi AI acceleratorIntel Gaudi 2 AI acceleratorIntel Gaudi 3AI acceleratorcodenamed Intel Gaudi 3AI acceleratorIntel Gaudi 2
6、 AI acceleratorAI Accelerator Advancements Available Now7nm 2x AI Compute 4x AI Compute 2xNetwork Bandwidth1.5xMemory Bandwidth20245nm Architected for Gen AIPerformance&ProductivityIncreased memory for LLM efficiency and cost effectivenessMassive,flexible on-chip networkingOpen standard vs.proprieta
7、ry InfiniBand64 Tensor Processor Cores(5th gen)8Matrix MathEngines128GB HBM capacity,3.7 TB/s B/W96MB SRAM,12.8 TB/s SRAM B/W24x 200 GbE industry-standard RoCE Ethernet portsPCIe 5 x 16Intel Gaudi 3 AI acceleratorI N T R O D U C I N G Designed for AI driving greater efficiency&performancePerformance
8、Scalable SystemsReference Architectures*Visuals for illustrative purposes,not actual systems.512-node Cluster(x4096 Accelerators)1 Node(x8 Accelerators)64-node Cluster(x512 Accelerators)1024-node Cluster(x8192 Accelerators)Inference-Lower latency,higher throughputTraining-Faster Time to train,larger
9、 model sizesPeak projected performance,memory capacity&B/W,networking scale-up/scale-out B/W Performance varies by use,configuration and other factors.Results may varyNetworking B/WFP8 ComputeMemory Capacity14.7 PF1024 GB8.4 TB/sNetworking B/WFP8 ComputeMemory Capacity940 PF65.5 TB76.8 TB/sNetworkin
10、g B/WFP8 ComputeMemory Capacity7.52 EF525.3 TB614 TB/sNetworking B/WFP8 ComputeMemory Capacity15 EF1 PB1.229 PB/sThe Challenge:Energy Consumption for GenAISource:Stanford HAI AI Report 2023 page 121Source:Stanford HAI AI Report 2023 page 120Parameters(billion)GPT-3GopherTrainingPower Consumption(MWh
11、)GPT-3Gopher5.31Human Life Avg,1 yearCar,avg incl fuel(lifetime)GPT-3GopherTraining CO2 equivalent emissions(tonnes)Data shown for model trainingInference60%Training40%Google ML EnergyConsumption*https:/ AI More Energy Efficient&SustainableOptimizeModelsOptimizeSoftwareModel CompressionAI&Analytics
12、ToolkitoneAPIOpen VINOKubernetesIntel Xeon,Gaudi,Core processor innovationData Center Liquid CoolingThe Future:A Sustainable Datacenter11Seamless Software migration across architecturesCommercial and technical supportCPU+Accelerator SecureSecureSystemsSystems-LedLedFederated LearningFraud DetectionS
13、ustainable AIInvestmentsInvestmentsUnconstrained SupplyCustomer engineering supportWorld Class FoundryOpen Software EcosystemOpen Software EcosystemRAS&SecurityDeveloper productivity for EnterpriseHardware AgnosticChoiceChoiceNetworking InvestmentsTime-to-market solutions at lower TCO today that mit
14、igates lost opportunity costCall to Action Lean in on the opportunities with AI and lets address the challenges together for a better,more sustainable AIPartner with OCP to help shape future open AI systems and standards -to bring open AI everywherePlease join us for our Systems session todayScaling
15、 Systems for Gen AI|Auditorium II|14:14-14:29Notice and Disclaimers 13Statements in this document that refer to future plans or expectations are forward-looking statements.These statements are based on current expectations and involve many risks and uncertainties that could cause actual results to d
16、iffer materially from those expressed or implied in such statements.For more information on the factors that could cause actual results to differ materially,see our most recent earnings release and SEC filings at .Allproduct plans and roadmaps are subject to change without notice.Performance varies
17、by use,configuration and other factors.Learn moreon thePerformance Index site.Intel technologies may require enabled hardware,software or service activation.Performance results are based on testing as of dates shown in configurations and may not reflect all publicly available updates.See backup for
18、configuration details.No product or component can be absolutely secure.Your costs and results may vary.Intel does not control or audit third-party data.You should consult other sources to evaluate accuracy.Code names are used by Intel to identify products,technologies,or services that are in develop
19、ment and not publiclyavailable.These are not commercial names and not intended to function as trademarks.Intel Corporation.Intel,the Intel logo,and other Intel marksare trademarks of Intel Corporation or its subsidiaries.Other names and brands may be claimed as the property of others.1415Comprehensive Portfolio that meets the needs of EveryoneAI ContinuumModel CreationTraining/Fine-TuningDeploymentEdge InferenceLocalized Inference(Client)Data PrepCLOUD&ENTERPRISECLIENT&WORKSTATIONEDGEAI ModelsDirectMLHugging FaceML applications