Hardware Datasheet · GPU Server
NVIDIA
NVIDIA DGX GB200 NVL72
Seventy-two Blackwell GPUs and thirty-six Grace CPUs as one liquid-cooled NVLink domain.
Overview
GB200 NVL72 is a rack, not a server. Seventy-two Blackwell GPUs and thirty-six Grace CPUs are connected as a single NVLink domain and delivered liquid-cooled as one unit, giving a coherent memory space far larger than any individual node can offer. It is the platform for models that do not fit inside an eight-GPU server, and it is a facility project as much as a hardware purchase, so the power density, floor loading and facility water need scoping before anything is ordered.
Specifications
| GPU / Accelerator | 72× NVIDIA Blackwell GPUs |
| CPU | 36× NVIDIA Grace CPUs |
| GPU Interconnect | 5th-generation NVLink, single 72-GPU domain |
| Cooling | Direct liquid cooling, facility water required |
| Scale-out fabric | InfiniBand or Ethernet, specified with the rack |
| Form Factor | Full rack, delivered and commissioned as one system |
| Site requirements | Power density, floor loading and facility water assessed before quote |
| NVLink Bandwidth | 130 TB/s aggregate rack-wide; 3.6 TB/s per Grace Blackwell Superchip |
| Total GPU Memory | 13.4 TB HBM3e, 576 TB/s aggregate bandwidth |
| Grace CPU Memory | 17 TB LPDDR5X, 14 TB/s bandwidth |
| CPU Cores | 2,592 Arm Neoverse V2 cores (36x Grace CPU) |
| Rack Power | Approximately 120kW |
| Compute Trays | 18x 1RU trays, each 2x Grace CPU + 4x Blackwell GPU |
| NVLink Switch Trays | 9x 1RU trays, 2 NVSwitches each (72 ports), 57.6 Tb/s per tray |
| Power Shelves | 8 shelves, 6x 5.5kW PSUs each (N+N redundant), 33kW per shelf |
| Compute Tray Networking | 4x ConnectX-7 single-port 400G OSFP; 2x BlueField-3 DPU dual-port 400G |
| Local Storage (per tray) | 4x 3.84TB E1.S NVMe (RAID 0) + 1x 1.92TB M.2 NVMe boot |
Typical Use Cases
- ·Frontier-scale model training
- ·Inference on models exceeding single-node memory
- ·Rack-scale AI factory build-outs
- ·SuperPOD-class deployments
Industries
Warranty & Support
Supplied through authorized channels with full manufacturer warranty. On-site, next-business-day support options available. Every system is configured, tested, and documented before delivery, with asset and warranty records provided for enterprise audit requirements.
Request a tailored quote
Configurations are tailored per engagement, so contact us for pricing and lead times.
sales@nexus-compute.com
+1 656 218 9096
nexus-compute.com
Specifications are indicative and configured to each engagement. All product names, logos, and trademarks are the property of their respective owners. Nexus Compute is an independent enterprise hardware supplier.