一个完全开放和协作的人工智能生态系统.pdf

1、Mark PapermasterCTO and EVPAMDA Fully Open and Collaborative AI Ecosystem20242025202620272028$45B$500BData Center AI Accelerator TAMInference Growing at 80%CAGRTrainingInference 202360%CAGRSource:AMD Internal Data 2024Performance/Watt andCapital Expense Space,Power,and CoolingOpen-Source Softwarevs

2、Proprietary StackGeneral Purposeand Custom Compute OptimizationAIChallengesOpen Accelerates AdoptionMobileAIInternetLinuxOpen Ecosystems Thrive ChoiceRapid Co-InnovationPortabilityProvenFlexibilityOpen HardwareOpen SoftwareOpen EcosystemDay-0 Supportfor Leading ModelsAcceleratedPace of InnovationBro

3、adeningEcosystem PartnershipsDeveloper FirstApproach to EnablementGPTRapidly Advancing Open Software Capabilities|Tightly Coupled Shared Memory with Massive BandwidthEfficient RDMA NetworkingSource:AMD Internal DataAI Foundation Models Require Rack ScaleRackData CenterSiliconServerNumber of GPUs1472

4、100,000+|Creating Choice and Open Innovationfor Scale-Up and Scale-Out NetworksAny CPU Any Accelerator Any NIC Any Switch|Creating Choice and Open Innovationfor Scale-Up and Scale-Out NetworksAny CPU Any Accelerator Any NIC Any SwitchESUN|Open Rack Wide(ORW)Available in 2026AMD“Helios”Optimized AI R

5、ack Solution|AI NICsConnected Directly to GPUs with UALinkEPYC CPUsInstinct GPUsInfinity Fabric Connected to EPYCUALink Over Internet|Confidential Computing ConsortiaAcceleratorsSmartNICsStorageTrustedBoundariesTrusted IOTDISP-Based EndpointsTDISP-Based EndpointsTDISP-Based Endpoints|October 2024Adv

6、isory FormationOne Year Later Advanced Performance Extensions(APX)Vector Instruction Set Extension(AVX10)Memory Safety(“ChkTag”)Flexible Return and Event Delivery(FRED)October 2025Results:Aligned x86 Ecosystemx86 Ecosystem Advisory Group Ad