MGX OCP - Meta 的下一代人工智能系统.pdf

《MGX OCP - Meta 的下一代人工智能系统.pdf》由会员分享，可在线阅读，更多相关《MGX OCP - Meta 的下一代人工智能系统.pdf（24页珍藏版）》请在三个皮匠报告上搜索。

1、Hao Shen,HW Engineer,MetaMatt Bowman,HW Engineer,MetaMGX OCP-Metas Next Gen AI SystemMGX OCP-Metas Next Gen AI SystemHao Shen,HW Engineer,MetaMatt Bowman,HW Engineer,MetaArtificial IntelligenceWhy MGX-OCP?Lots of custom hardwareLots of leveraged hardwareGrand Teton(H100)Why MGX-OCP?Lots of custom ha

2、rdwareLots of leveraged hardwareGrand Teton(H100)Catalina(GB200)Why MGX-OCP?Lots of custom hardwareLots of leveraged hardwareMGX-OCP(GB300)Catalina(GB200)Grand Teton(H100)Codesign with NV for this reference Enable larger validation pool Proactively find and fix bugs early Common supply chain itemsMG

3、X-OCP Collaboration OCP network cards that provide out of band connections through network controller sideband interface(NSCI)Data Center Secure Control Module(DC-SCM)for management Leak management Power Distribution Board heatsink Interposer that glues it all togetherThe Changes?MGX-OCPGB300 Module

4、sHMCCX8 IO BoardPDBOSFP Carrier boardFront End NICsE1.S NVME BackplaneFansCold Plate LoopBoot DriveInterposerE1.S Data DrivesBack End NetworkDC-SCM 2.0Leak Sensing and ControlMGX-OCPGB300 ModulesHMCCX8 IO BoardPDBOSFP Carrier boardFront End NICsE1.S NVME BackplaneFansCold Plate LoopBoot DriveInterpo

5、serE1.S Data DrivesBack End NetworkDC-SCM 2.0Leak Sensing and ControlMGX-OCPGB300 ModulesHMCCX8 IO BoardPDBOSFP Carrier boardFront End NICsE1.S NVME BackplaneFansCold Plate LoopBoot DriveInterposerE1.S Data DrivesBack End NetworkDC-SCM 2.0Leak Sensing and ControlMGX-OCPGB300 ModulesHMCCX8 IO BoardPD

6、BOSFP Carrier boardFront End NICsE1.S NVME BackplaneFansCold Plate LoopBoot DriveInterposerE1.S Data DrivesBack End NetworkDC-SCM 2.0Leak Sensing and ControlMGX-OCPGB300 ModulesHMCCX8 IO BoardPDBOSFP Carrier boardFront End NICsE1.S NVME BackplaneFansCold Plate LoopBoot DriveInterposerE1.S Data Drive