GB200硬件体系结构——组件供应链和M

Key Points 1. GB200 Hardware Architecture and Component Supply Chain * GB200 Hardware Architecture: The GB200 features a superior hardware architecture with significant performance improvements. However, its deployment complexity has increased dramatically. The system includes various components such as NVLink backplanes, NVLink 36, NVLink 72, NVLink 576, PCIe retimers, switches, optics, DSP, PCB, and liquid cooling. * GB200 Form Factors: There are four main form factors for the GB200 rack, each with customization options. These include NVL72, NVL36x2, NVL36x2 (Ariel), and B200 NVL72/NVL36x2 (Milan). * Power Budget: Each NVL72 rack requires approximately 120 kW, while NVL36x2 racks require 132 kW. This high power demand necessitates liquid cooling for the GB200. * Component Supply Chain: The supply chain for the GB200 has been redesigned to accommodate the new hardware architecture and power requirements. This includes changes to the component BOM, liquid cooling architecture, and deployment complexity. 2. NVLink Interconnect * NVLink Topology: The GB200 uses NVLink for accelerator interconnect, providing high-speed communication between GPUs. NVL72 maintains the same NVLink topology as HGX Hopper/Blackwell, while NVL36x2 requires two NVSwitch jumps to communicate with GPUs in adjacent racks. * NVLink Cost: The NVLink interconnect system is expensive due to the use of active copper cables (ACC) and NVSwitch ASICs. The cost of NVLink backplanes and cables is significantly higher than previous generations. 3. Backend Network * Backend Network Options: The GB200 supports various backend network options, including InfiniBand, RoCE Ethernet, and Spectrum-X Ethernet. The initial shipments will use ConnectX-7 NICs, with future options including ConnectX-8 and Spectrum-X Ultra 800G. * Switch Options: Customers can choose from different switch options, such as QM9700, QM3200, and XSN5600. The switch selection depends on the specific requirements of the customer's network architecture. 4. Frontend Network * Frontend Network Options: The GB200 supports various frontend network options, including standard Ethernet and BlueField-3. The initial shipments will use ConnectX-7 NICs, with future options including ConnectX-8 and Spectrum-X Ultra 800G. * BlueField-3 Usage: BlueField-3 usage is expected to be limited, primarily used by Oracle for network virtualization purposes. 5. Liquid Cooling * Liquid Cooling Solutions: The GB200 requires liquid cooling due to its high power demand. There are two main liquid cooling solutions: Liquid to Air (L2A) and Liquid to Liquid (L2L). L2L is expected to be the mainstream solution due to its energy efficiency and cost advantages. * Liquid Cooling Components: The liquid cooling components include cold plates, CDMs, CDUs, and QDs. The supply chain for these components is complex and involves various suppliers. 6. Power Supply * Power Supply Network: The GB200 uses a centralized power supply network at the rack level, with power distribution units (PDUs) and busbars. This design increases power efficiency compared to traditional node-level PSUs. * Power Components: The power supply components include PSUs, busbars, and VRMs. The supply chain for these components is diverse, with various suppliers participating. 7. BMC * BMC Content: The GB200 has a higher BMC content compared to HGX due to the increased CPU-to-GPU ratio. The BMC content is driven by the need for remote monitoring and management of the system. * BMC Market: The BMC market is expected to grow due to the increasing demand for AI servers and other data center equipment. 8. Mechanical Components * Chassis and Rails: The GB200 requires high-specification chassis and rails due to its size and weight. The supply chain for these components includes various suppliers, with King Slide and Nan Juen leading in the 5U+ rail kit market. * Chassis Content: The chassis content for the GB200 is higher than traditional servers due to the larger size and higher power requirements. 9. OEM/ODM Mapping * OEM/ODM Details: The GB200 deployment complexity has increased, leading to higher requirements for OEM/ODM expertise. The supply chain for OEM/ODM services is diverse, with various players participating. 10. GB200 Component and Supply Chain Model * GB200 Component and Supply Chain Model: The GB200 component and supply chain model provides detailed information about the components, their prices, and the supply chain landscape. This model helps in understanding the cost structure and supplier landscape for the GB200.