Aivres公司展示了采用NVIDIA GPU的OCP解决方案助力新一代人工智能发展.pdf

1、OCP Solutions with NVIDIA GPUs for New Gen AIJames ZouR&D Fellow,AivrestttAivres OCP SolutionsNVIDIA H200,NVIDIA RTX PRO 6000KR4268 4U AC/KR6268 6U AC SystemSupports 8 FHFL PCIe GPU Cards,up to 2 additional double-width PCIe GPUsNVIDIA HGX B200/B300 8-GPUKR9288 10U AC/KR5288 5U DLC SystemHigh-Perfor

2、mance Accelerated ServerSupports 2 Intel Xeon 6/AMD EPYC CPUs up to 500W TDPNVIDIA GB200/GB300 NVL72KRS8000/KRS8500 Rack System Liquid-Cooled Rack-Scale ArchitectureAiVRES Full Mode MonitoringAiVRES Diagnosis AutomationTrillion Parameters AI Model(More Than 1T):GPT-4 1.8T AI ModelVery Large AI Model

3、(10B 1T):LLaMA 70B,and 405B AI ModelLarge AI Model(Fewer Than 10B):LLaMA 7B AI Model2OCP Full Stack Solution Based on NVIDIA GB200/GB300 NVL72OCP OAM Module SolutionBased on NVIDIA HGXB200/B300PCIe Card GPU SolutionBased on NVIDIA H200,RTX PRO 6000 or B300AOCP AI Solution Based on NVIDIA GPUs3PCIe C

4、ard GPU SolutionBased on H200,NVIDIA RTX PRO 6000 and B300AOCP OAM Module Solution Based on NVIDIA HGX B200/B300OCP Full Stack Solution Based on NVIDIA GB200/GB300 NVL72System CostSystem ComplexityManufacturing DifficultySystem Service DifficultyPerformancePower EfficiencyDensityKRS8500 OCP Full Sta

5、ck AI Solution on NVIDIA GB300 NVL724 72 NVIDIA Blackwell GPUs Networking system Storage system Liquid-cooling system Hardware monitoring and management Customization of networking and storage system Platform setup for AI development AI development use cases sharing Professional deployment supportNe

6、tworking CustomizationKRS8500 OCP Full Stack AI Solution Speed BandwidthStabilityScalabilityTopologyComputing NetworkUse blue-field DPU to offload N-S traffic and accelerate computational efficiency.Use the converged network to simplify the interface.Storage NetworkIntegrate server management and sw