Address
304 North Cardinal St.
Dorchester Center, MA 02124
Work Hours
Monday to Friday: 7AM - 7PM
Weekend: 10AM - 5PM
Plain-language- GPU Server Hardware Composition and Disassembly: What Sales Need to Know
GPU Server Overview and Exterior
ASUS (vendor) HGX H100 server GPU module and front tray (partially pulled):
Frontal view / front panel of some vendor’s GPU servers:
Rear view / rear panel of some vendors’ air-cooled GPU servers:
Rear view / rear panel of a vendor’s liquid-cooled GPU server:
GPU Server Modules and Components
The following diagram show the major modules of a GPU server:
The following figure shows the individual parts within a GPU server:
Among the parts of a GPU server, the two core modules (excluding the chassis) are:
1. The GPU node (the GPU module board and attached components).
2. The Front / CPU compute node (the CPU compute board and attached components).
GPU Module and Composition
GPU module overview:
key components of a GPU module:
Key components of a GPU module:
GPU module board (UBB, Unit Base Board): Hosts multiple GPUs intergrated into a matrix platform, providing high-speed data exchange among GPUs and between GPUs and CPUs. OAM GPU Module: 是An Open Accelerator Module standard-based GPU form factor (e.g., SXM A100 GPUs) that can be installed on the UBB. NVSwitch chips: Enable ultr-high-speed communication betweeen multiple GPUs. GPU heatsink / cooler: Provides thermal dissipation for the GPUs.
CPU Compute Node (Front Head) and Components
CPU compute node (front head) includes the following parts:
1-CPU compute node chassis cover: Mount on the CPU compute node chassis to protect internal components. 2-Storage controller card: Provides RAID support for SAS/SATA drives, supports RAID configuration and expansion, firmware updates, and remote setup. 3-Riser / Riser card: Adapter card that allows PCIe cards to be mounted into the server via riser connections. 4-Supercapacitor mounting bracket: Used to secure supercapacitors within the chassis. 5-Server management module: Provides a variety of I/O interfaces and out-of-band management (BMC/HDM) functionality. 6-OCP adapter module: Used to install OCP netword cars (designed according to Open Compute Project specifications). 7-Air shroud / airflow guide: Provides cooling airflow channels for CPU heatsinks and memory, and accommodates supercapacitor placement. 8-CPU heatsink shroud: Provides cooling for the CPU. 9-Memory (RAM): temporarily stores compute data for the CPU and exchanges data with persistent storage, supports DDR5 memory, RDIMM or LRDIMM. 10-CPU: Intergrates memory controller and PCIe controller, supplying the server’s primary compute capability. 11-Standard PCIe nework: A NIC that installs into standard PCIe slots. 12-Rear drive cage: Supports expansion of rear-mounted drives. 13-Network card adapter: Available in 4-slot and 8-slot variants to accommodate different numbers of NICs. 14- OCP network card: Only supports installation into the OCP adapter module. 15-Busbar: Connects power between the power conversion board and the PCIe switch adapter board. 16-Power conversion board: Supplies power to GPU nodes and reports power status to the motherboard. 17-Encryption / security module: Provides cryptographic services to enhance data security. 18-M.2 SSD card: Provides storage media for the server. 19-Supercapacitor: Supplies power to storage controller flash in the event of unexpected power loss to protect data. 20-CPU compute node power module: Provides power conversion for the CPU compute node. Supports hot-swap and 1+1 redundancy. 21-GPU power module: Provides power conversion for GPU nodes, fans, front drive bays, and NIC adapter modules. Support hot-swap and typically 3+3 redundancy. 22-PCIe Switch adapter board: Used to extend PCIe signals; houses PCIe switch to efficiently interconnect GPUs with storage and NICs. 23-Motherboard: One of the server’s most critical components; hosts CPU, memory, BIOS, PCIe slots, and other fundamental system components. 24-Compute node chassis: 将The enclosure that houses all compute node components.
