1/2
NVIDIA GB300 NVL72
GIGAPOD AI DLC POD Scale Solution
- Built with NVIDIA GB300 Grace™ Blackwell Ultra Superchips
- 16 x NVIDIA GB300 NVL72 racks
- 1152 x NVIDIA Blackwell Ultra GPUs and 576 x NVIDIA Grace™ CPUs
- 272 TB of LPDDR5X ECC memory
- 320 TB HBM3E GPU memory
- Up to 140 kW rack power consumption
- Reserved space for InfiniBand / Ethernet switches
- Integrated G-REX cluster-level DLC power and cooling management kit
- Compatible with GIGABYTE POD Manager software for cluster management / orchestration
- Supports flexible storage with GIGABYTE storage servers and validated partner solutions
Configuration
| PAC2 Series | |
|---|---|
| Cooling Type | Direct Liquid Cooling |
| Racks per Scalable Unit (Compute + Management) | 16+8 |
| Rack Height | 48U |
| Model | NVIDIA GB300 NVL72 |
| Compute Tray per Rack | 18 x GIGABYTE XN15-CB0-L01 |
| CPU | 576 x NVIDIA Grace™ CPUs |
| GPU | 1152 x NVIDIA Blackwell Ultra GPUs |
| Switch | Flexible validated InfiniBand / Ethernet switch configurations |
| Power Shelves per Rack | 6 x 1U 33 kW power shelves |
| Rack Power Consumption | Up to 140 kW |
| CDU | In-Rack / In-Row |
| CDU Specifications | In-Rack Liquid-to-Liquid CDU: Cooling Capacity: 200 kW (4°C approach temperature) Primary Loop Connections: 1" FD83 QDC couplings (inlet/outlet) Input Power: 1P / AC 220-240 V / 50-60 Hz 1+1 redundant power supplies and pumps In-Row Liquid-to-Liquid CDU: Cooling Capacity: 1350 kW (4°C approach temperature) Primary Loop Connections: 5" clamp interface (inlet/outlet) Input Power: 2 x IEC 60309 100A (4P5W, 6h plug), AC 380 V ±10% / 3P / 60 Hz 1+1 redundant pumps and redundant A/B power feeds [Note] The Facility Water System (FWS) loop must meet the water quality values specified in ASHRAE Liquid Cooling Guidelines for Datacom Equipment Centers, 2nd ed., Ch. 5, as reproduced in the ASHRAE TC 9.9 white paper Water-Cooled Servers: Common Designs, Components, and Processes (2019), Table 1. |
| Management | Integrated G-REX DLC power and cooling management kit: - N120-LS0 management node - Real-time cluster-level monitoring and control - Supports Modbus, Redfish, and REST APIs for seamless integration with CDUs and RDHx Optional GIGABYTE POD Management (GPM) software suite: - Essential single-node BMC for real-time hardware health and remote system oversight - Unified cluster-scale management and deep telemetry for high-density deployments - Total POD-level orchestration for infrastructure provisioning and environmental monitoring |
| Storage Solutions | Flexible validated partner solutions with GIGABYTE storage servers |
| Cooling Type |
| Racks per Scalable Unit (Compute + Management) |
| Rack Height |
| Model |
| Compute Tray per Rack |
| CPU |
| GPU |
| Switch |
| Power Shelves per Rack |
| Rack Power Consumption |
| CDU |
| CDU Specifications |
| Management |
| Storage Solutions |
[#1] All materials provided herein are for reference only. GIGABYTE reserves the right to modify or revise the content at any time without prior notice.
[#2] Advertised performance is based on maximum theoretical values as specified by the respective chipset vendors or standards organizations. Actual performance may vary depending on system configuration.
[#3] All trademarks and logos are the property of their respective owners.
