通過可持續計算實現智能.pdf

《通過可持續計算實現智能.pdf》由會員分享，可在線閱讀，更多相關《通過可持續計算實現智能.pdf（29頁珍藏版）》請在三個皮匠報告上搜索。

1、Enabling IntelligenceThrough Sustainable ComputingUnrestricted|AI Hardware Summit|Siemens 2024|Siemens Digital Industries SoftwaresustainabilitynounThe ability to be maintained at a certain rate or level:the sustainability of economic growth the long-term sustainability of the project“Avoidance of t

2、he depletion of natural resources to maintain an ecological balance:the pursuit of global environmental sustainability the ecological sustainability of the planetUnrestricted|AI Hardware Summit|Siemens 2024|Siemens Digital Industries SoftwareThree pillars of sustainabilityUnrestricted|AI Hardware Su

3、mmit|Siemens 2024|Siemens Digital Industries SoftwareTechnologicalEnsuring the product can be built and is fit for purposePowerPerformanceetcEconomicCan it be built in a way that is profitableCost of manufactureCost of ownershipetcEnvironmentalMinimizing the negative impact on the environmentEnergy

4、to manufactureMaterialsetcNeed to keep all three in balance to build successful businesses,it is bigger than just environmental sustainabilityAI adoption trending upwards,with GenAI usage doubling since 2023Unrestricted|AI Hardware Summit|Siemens 2024|Siemens Digital Industries SoftwareGrowth will c

5、ontinue to accelerate as more use-cases come online and larger models expand applicabilitySource:McKinseyAI model size increase continues to rapidly increaseUnrestricted|AI Hardware Summit|Siemens 2024|Siemens Digital Industries SoftwareAs AI model size grows so does its capability and demands on su

6、stainabilityMore complex technology needsHigher power demandsGreater resource utilizationSource:EpochThe AI growth drives compute demand,boosting the semiconductor industryUnrestricted|AI Hardware Summit|Siemens 2024|Siemens Digital Industries Software22%CAGR Source:GartnerComplexity is compoundingH

7、ardware and software complexity is linked and growing at an exponential rateSiP era2020Transistor eraGate eraASIC eraSoC era1980199020002010HeuristicsReinforcementlearningGraph analyticsResourcesSW ComplexityHW ComplexityIntelligentSystems Era2030Unrestricted|AI Hardware Summit|Siemens 2024|Siemens

8、Digital Industries SoftwareAI and sustainability are on a collision courseUnrestricted|AI Hardware Summit|Siemens 2024|Siemens Digital Industries SoftwareEnergy demand growth significantly outpaced supply growthHitting an energy wall by 2040Powerful new AI models requires more energy consumption,the

9、re is no way to get there without energy breakthrough.-Sam Altman,OpenAIUnrestricted|AI Hardware Summit|Siemens 2024|Siemens Digital Industries SoftwareOn the current path AI applications will be“Energy Bound”Unrestricted|AI Hardware Summit|Siemens 2024|Siemens Digital Industries SoftwarePower limit

10、s technology adoption,limits industry growth,and impacts corporate sustainability goalsPower InfrastructureDevice physicsChatGPT consumes a lot of energy in the process,up to 25 times more than a Google search.Additionally,a lot of water is also used in cooling for the servers that run all that soft

11、ware.Per conversation of about 20 to 50 queries,half a liter of water evaporates a small bottle,in other words-Brussels TimesPerformance per Watt is now the limiting factorUnrestricted|AI Hardware Summit|Siemens 2024|Siemens Digital Industries SoftwareCompute performance is growing slower than power

12、 demand due to slowing node scalingDouble impact as increased compute power requires increased cooling power and resourcesDemandSupplyGapNvidias upcoming Blackwell generation boosts power consumption even further,with the B200 consuming up to 1,200W,and the GB200(which combines two B200 GPUs and one

13、 Grace CPU)expected to consume 2,700W.This represents up to a 300%increase in power consumption across one generation of GPUs with AI systems increasing power consumption at a higher rate.-ForbesPerformance per watt scalingScope of potential solutionsUnrestricted|AI Hardware Summit|Siemens 2024|Siem

14、ens Digital Industries SoftwareDeviceDeliver more efficient semiconductor devicesBetter utilization of existing technology solutionsProductReduce buffer and waste in design architectureImproved functional allocation between HW/SW,Electrical/MechanicalLifecycle(concept end-of-life)Impact of supply ch

15、ain membersManufacturing resource utilizationRecyclability and longevityAddressing energy challenges while continuing growth of AIUnrestricted|AI Hardware Summit|Siemens 2024|Siemens Digital Industries SoftwareWorkload Optimized SiliconWorkload-optimization leads to greater energy efficiencyUnrestri

16、cted|AI Hardware Summit|Siemens 2024|Siemens Digital Industries SoftwareCPU(AMD EPYC 7763)GPU(A100)1x60 x218xBroader Range of AI AcceleratorsPerformance per watt Application-SpecificAlgorithm-SpecificDedicatedTPU(Google TPU V4)Evolving compute infrastructure:data center to edge for AI applicationsUn

17、restricted|AI Hardware Summit|Siemens 2024|Siemens Digital Industries SoftwareData CenterEdgeLevel of Optimization for AI WorkloadsTransformation is needed on the edgeUnrestricted|AI Hardware Summit|Siemens 2024|Siemens Digital Industries SoftwareEnvironmental Footprint(Source:World Economic Forum)2

18、0%80%Training:Centralized computing in the data centersInference:Decentralized processing on the edge and KEY to a sustainable AI compute infrastructureHowever,designing workload-optimized hardware has challengesUnrestricted|AI Hardware Summit|Siemens 2024|Siemens Digital Industries SoftwareUnique d

19、esignrequirementsComplex&lengthy processesLimitedreusabilityLong development cycleHigh development costUsing traditional approaches optimizing devices for their unique workloads has increased overall business riskOptimization traditionally favors centralized resourcesUnrestricted|AI Hardware Summit|

20、Siemens 2024|Siemens Digital Industries SoftwareData CenterLarge-scale deployment offset high dev costsSingle owner of cost easily justify investmentScale of deploymentData centerEdgeEdgeDiversity of applications complicates finding commonality to scaleHigh dev cost make investment harder to justify

21、 Business Investment ThresholdMaking workload-optimized hardware on the edge viableUnrestricted|AI Hardware Summit|Siemens 2024|Siemens Digital Industries SoftwareScale of deploymentData centerEdgeBusiness Investment ThresholdAutomation is key to shifting the curve up,reducing development cycle,cost

22、,and achieving viability.Higher abstraction accelerates product developmentUnrestricted|AI Hardware Summit|Siemens 2024|Siemens Digital Industries SoftwareAlgorithm LevelAlgorithm in SystemC/CPrompt LevelDesign requirements in textApplication LevelAl frameworks in PythonRegister Transfer LevelHardwa

23、re representation in HDLRTL SynthesisHigh-Level SynthesisHigh-Level Synthesis for AIYearsMonthsWeeksDaysHigh-Level Synthesis with GenAICatapult AI NN:automation from AI network to semiconductor implementationUnrestricted|AI Hardware Summit|Siemens 2024|Siemens Digital Industries SoftwareAutomated Tr

24、anslation from High-level AI FrameworkOptimized Design Workflow on the Application LevelPythonTrained neural networkC+Inference engineRTLRight-sized AI acceleratorhls4mlHigh-Level SynthesisPPAArchitecture exploration for optimized PPASupport both ASIC and FPGAWakeword inference:a new use case unlock

25、ed with Catapult AI NNUnrestricted|AI Hardware Summit|Siemens 2024|Siemens Digital Industries Software90%Reduction in energy per inference15XFaster execution timeConfigurationSpeedmsecsPowermWAreaSq micronEnergy/InfuJRISC-V Software285189393,52133,577Software+Accelerators18203722,5213,683Recognizing

26、 your top careaboutsUnrestricted|AI Hardware Summit|Siemens 2024|Siemens Digital Industries Software012345678910Perf/W/$TCOReliabilityCost(Yield)MemoryPerformanceAI ComputePowerNetworkingPowerCPU PowerOther PowerCloud AI(large-scale training,inference)Total datacenter power limits model size Network

27、 power contribution is significant Endpoint AI(personal assistants,games,copilot)Trade-off between user experience and battery lifeEmbedded/Edge AI(automotive,robotics,medical)Total power envelope highly constrainedCloud AIEndpoint AIEmbedded AIComponents of powerPPA optimized C-to-GDS solutionLifec

28、ycle Management to deliver best perf/TCO and addressing challenges throughout the lifecycleDesignProductionIn-LifeWorkload optimized DesignEmbedded Analytics over product lifecycleAdvanced node/3DIC supportUnrestricted|AI Hardware Summit|Siemens 2024|Siemens Digital Industries SoftwareBut Semiconduc

29、tors are just part of a more complex systemUnrestricted|AI Hardware Summit|Siemens 2024|Siemens Digital Industries SoftwareSolving the challenges for Software Defined ProductsUnrestricted|AI Hardware Summit|Siemens 2024|Siemens Digital Industries SoftwareRequirementsManagementSimulation DrivenDecisi

30、onsMulti-DomainProductEngineeringSimulation DrivenVerificationSupplierCollaborationDigitalManufacturingCertification&QualificationElectronicsElectricalMechanicalSoftwareCONTINUOUS PRODUCT SIMULATION AND OPTIMIZATIONCLOSED-LOOP DIGITAL TWINManufacturingOperations&ExecutionRAMSSystemsArchitectureConce

31、ptualDesignAsset Operations and SustainmentWhich requires a Comprehensive Digital TwinUnrestricted|AI Hardware Summit|Siemens 2024|Siemens Digital Industries SoftwareUnderstand design and manufacturing considerations early in the design processPerform design trade-offs with visibility into the impli

32、cations of any changeMake decisions with a more holistic view of the product lifecycleDigital Twin ProductDigital Twin ProductionReal ProductionReal ProductOptimizing edge AI Infrastructure drives growth and scales sustainabilityUnrestricted|AI Hardware Summit|Siemens 2024|Siemens Digital Industries

33、 SoftwareSiemens offers solutions enabling a sustainable computing futureEnergyEfficiencyResourceEfficiencyInfrastructureEfficiencyScalability of systemsSustainable manufacturingLifecycle solutionsContactPublished by Siemens 2024Ankur GuptaSVP and GM Digital Design Creation Platform(DDCP)Siemens EDAUnrestricted|AI Hardware Summit|Siemens 2024|Siemens Digital Industries Software