Wiwynn 助力构建人工智能的未来——加速向人工智能驱动型基础设施转型（Wiwynn 出品）.pdf

1、Alfred ChaoLine of Business HeadWiwynn CorporationBuilding the Future of AIWiwynn Accelerates the Shift to AI-Driven InfrastructureShifts in the Server IndustryAI Demands Drive Technical Evolution Emerging Players Create Market Opportunities Service Scope Enhancement Explosive Growth of Computing Po

2、wer in the AI Era4Growth of AI Model ScaleSource:https:/cloudswit.ch/blogs/ai-data-center-network-traffic/GPT-1BERT LargeXLNetMegatron-NLGGPT-2 1.5BMicrosoft T-NLGGPT3-175BFalcon 180BMT NLG 530BGPT-MoE-1.8TChinchillaBLOOMPaLMWav2Vec 2.0MoCo ResNet50ELMoResNeXtDenseNet201XceptionInceptionV3ResnetVGG-

3、19AlexNetBefore Transformer8x/2yrsTransformer256x/2yrs201220142016201820202022202410010,0001,000,000100,000,00010,00,000,000Training Compute(PetaFLOPs)The Transformer architecture enabled AI models to scale up to billions of parametersSparked a dramatic increase in required computing power.201220142

4、0162018202020222024Rising Power Consumption in Chip Design5202620282030Power Requirement Trend for CPU/GPUSource:Companies and Morgan Stanley ResearchSky LakeEagle StreamMountain StreamBirch StreamINTELHopperAmperePascalVoltaBlackwellNVIDIANaplesRomeMilanGeonaTurinAMD0600Thermal Design Power(W)36003

5、0012002400MI210MI250MI300AMI325MI350Thermal Challenges Drive Advanced Cooling Innovation6HeatsinkCooling Solutions Integrated with IC Package2024202620282030Cold Plate&Immersion Cooling Air Cooling1.2x/2yrsAdvanced Cooling2x/2yrsFuture Cooling3x/2yrs01,0002,0003,0005,0004,000Cooling Capacity(W)20122

6、01420162018202020222024AI Demands Drive Technical Evolution Cross IndustriesMECHANICALWider Rack DesignInnovative Material Reduction DesignHigh Flexibility in Customization&IntegrationPOWERHigh Power Density&EfficiencyHVDCModular Power ManagementSafety,Redundancy&CompatibilityMore Layer,High-Speed,H