FusionPoD for AI Rack-Scale Liquid-Cooled Server

Fusion PoD for AI

Introduction

 

The FusionPoD for Al rack-scale liquid-cooled server adopts the leading integrated cabinet design. It provides a 105 kW high-power centralized power supply system and eight nodes with each supporting eight GPUs. In addition, the cabinet uses the power supply and liquid cooling bus blind mating, 2N/N+N redundancy backup, and native liquid cooling, which achieves extreme energy efficiency and ultra-low PUE. The rack-scale liquid-cooled server supports automatic robot O&M and unmanned intelligent detection and O&M.
FusionPoD for Al features high computing, performance, energy efficiency, reliability, and bandwidth, integrated cabinet delivery, and simplified O&M. It is ideal for AI scenarios such as training and inference of large and super large models.

 

 

Highlights

Compatibility with Diverse Computing and High-Density Integration of 64 GPUs
Diverse computing in the same architecture
High-density deployment of eight GPUs per node and eight nodes per cabinet 105 kW power supply and heat dissipation per cabinet
Flexible Ethernet and IB switches in the cabinet, up to 12 PCIe 5.0 x16 expansion slots, and cluster networking
Native Liquid-Cooling Design, Reducing TCO by 15% in Five Years
Multi-level liquid cooling leakage protection and detection for components, nodes, and cabinets
Water and electricity isolation and proactive pipe shutdown in case of liquid leakage
Highly reliable liquid-cooling components, providing a life span of 10+ years, with TÜV Rheinland liquid cooling leakage-proof certification
Green and low-carbon, 100% liquid cooling, and pPUE per cabinet as low as 1.06; integrated cabinet delivery, improving deployment efficiency by 10 times
Full Blind Mating of Power, Network, and Liquid, Supporting Plug and Play

Integrated cabinet deployment, improving rollout efficiency by 10 times
Nodes or GPU modules that are installed and removed independently, avoiding operations on power, network, or liquid during maintenance

Cabinet: FusionPoD for AI

Technical Specifications

- Environmental Specifications
• Operating temperature: 5°C to 40°C (41°F to 104°F), compliant with ASHRAE Classes A1 to A3
• Storage temperature (within three months): –30°C to +60°C (–22°F to +140°F)
• Storage temperature (within six months): –15°C to +45°C (5°F to 113°F)
• Storage temperature (within one year): –10°C to +35°C (14°F to 95°F)
• Maximum rate of temperature change: 20°C (36°F) per hour and 5°C (9°F) per 15 minutes

- Relative Humidity (No Condensing)
• Operating humidity: 8% to 90%
• Storage humidity (within three months): 8% to 85%
• Storage humidity (within six months): 8% to 80%
• Storage humidity (within one year): 20% to 75%
• Maximum rate of humidity change: 20% per hour

- Available Space
• Node area: 8 x 4U server nodes
• Switch area: 5 x 1U general-purpose switches (rear ports) or 4 x 1U general-purpose switches (front
ports)
• PSU area: 2 x 3U power shelves and 1 x 2U AC PDU
- Server Node:  8 x 4U server nodes
- pPUE:  ≥ 1.06
- Management Module
• Cabinet management module: 2 x 10GE ports, 1 x GE port, and 1 x RS485 serial port
• PSU, sensor, and asset management in the cabinet, and cabinet-level liquid leakage monitoring

- Switch Slot
• Up to 5 x slots for switches
• 5 x 1U general-purpose switches (rear ports) or 4 x 1U general-purpose switches (front ports)
• 1 x GE/10GE/25GE/100GE switch or 200G/400G IB switch supported by each slot
- Power Supply and Distribution

• 2N power supply system
• Nominal voltage output: 48 V; actual output voltage: 54.5 V
• 36 x PSUs (3 kW per PSU)
• PSUs in N+M redundancy (1 ≤ M ≤ N)
• Three-phase AC power supply input (6+6); voltage range: 346 V AC to 415 V AC
- Certification: CE, CB, and RoHS

- Cabinet Size (H x W x D)
• Optical cables: 2250 mm x 600 mm x 1200 mm (88.58 in. x 23.62 in. x 47.24 in.)
• Electrical cables: 2250 mm x 600 mm x 1300 mm (88.58 in. x 23.62 in. x 51.18 in.)
• Note: Height includes height of casters. When a RDHx is included, depth increases by 125 mm (4.92 in.)

- Weight in Full Configuration

• Fully-loaded cabinet: 1300 kg (2866.01 lb)
 

Node: GN560E V7

Technical Specifications

- Environmental Specifications

2 x 4th Gen Intel® Xeon® Scalable processors (Intel Eagle Stream Platform all series Sapphire
Rapids CPUs), up to 350 W TDP per processor
- Chipset: Emmitsburg PCH

- Memory

32 x 4800 MT/s DDR5 RDIMMs
- Local Storage
Hot-swappable and flexible drive configurations:
Up to 8 x 2.5" front NVMe drives + 2 x 2.5" front SAS or SATA drives
- RAID
Hot-swappable and flexible drive configurations:
Up to 8 x 2.5" front NVMe drives + 2 x 2.5" front SAS or SATA drives
- GPU Module
8 x GPU modules
- Network
Multiple network expansion capabilities: IB, Ethernet, and FC
Up to 12 x standard PCIe NICs, which can be configured as required
- PCIe Expansion
Up to 13 x standard PCIe slots
- Fan Module: 10 x fan modules in N+1 redundancy

- Cold Plate

Liquid cooling: CPUs + VRD + HGX 8-GPU modules

Liquid cooling: CPUs + memory modules + VRD + HGX 8-GPU modules

- Power Supply
54.5 V centralized power supply in a cabinet
- Management
The iBMC chip integrates one dedicated management GE network port, providing comprehensive management features such as fault diagnosis, automatic O&M, and hardware security hardening.
The iBMC supports standard interfaces such as Redfish, SNMP, and IPMI 2.0, provides a remote management interface based on HTML5/VNC KVM, and supports out-of-band management functions such as monitoring, diagnosis, configuration, Agentless, and remote control for simplified management.
OS: FusionOS, RHEL, Rocky Linux, Ubuntu, and SUSE

- Security: Administrator password, TPM 2.0, and secure boot

- Operating Temperature

Operating temperature: 5°C to 40°C (41°F to 104°F), compliant with ASHRAE Classes A1 to A3

- Certification: CE, CB, and RoHS

- Dimensions (H x W x D): 175 mm x 536 mm x 1006.4 mm (6.89 in. x 21.10 in. x 39.62 in.)
- Weight: 80 kg (176.37 lb)