Brand: NVIDIA | Category: Servers
SKU: NVID-MGXGB200NVL72 | Part #: MGX-GB200-NVL72 | MPN: MGX-GB200-NVL72
Contact for Pricing — Request a Quote
The NVIDIA MGX GB200 NVL72 Reference Server Platform from NVIDIA is enterprise-grade Servers hardware built for data centers. A rack-scale 72-GPU Blackwell AI supercomputer delivering 1.4 exaFLOPS of FP4 AI compute in a single NVLink-unified system. Available in brand new condition. Available from Omnixon Global — contact our team to request a quote.
| Brand | NVIDIA |
| Category | Servers |
| SKU | NVID-MGXGB200NVL72 |
| Part Number | MGX-GB200-NVL72 |
| Condition | New |
| Platform Architecture | NVIDIA MGX GB200 NVL72 — rack-scale reference design |
| GPU Count per Rack | 72× NVIDIA Blackwell B200 GPUs |
| CPU Count per Rack | 36× NVIDIA Grace CPUs (72 Arm Neoverse V2 cores each) |
| Superchip Configuration | 36× Grace Blackwell Superchips (1× Grace + 2× B200 per Superchip) |
| AI Compute (FP4) | 1.4 ExaFLOPS |
| AI Compute (FP8) | 720 PetaFLOPS |
| HBM3e Memory per GPU | 192 GB HBM3e |
| Total HBM3e Memory (Rack) | 13,824 GB (≈ 13.5 TB) aggregate GPU memory |
| GPU Memory Bandwidth per GPU | 8 TB/s |
| CPU-to-GPU Interconnect | NVLink-C2C at 900 GB/s bidirectional per Superchip |
| NVLink Switch All-to-All Bandwidth | 1.8 TB/s bisection bandwidth across all 72 GPUs |
| NVLink Generation | 5th Generation NVLink with NVLink Switch System |
| External Network Connectivity | Up to 400 Gb/s InfiniBand NDR or 400 Gb/s Ethernet per NVLink Switch tray |
| System Memory (CPU LPDDR5X) | 480 GB LPDDR5X per Grace CPU node (total ~17.3 TB across 36 nodes) |
| Storage Interface | NVMe SSDs via PCIe Gen 5 on each compute tray; network-attached storage integration via external fabric |
| Thermal Solution | Direct Liquid Cooling (DLC), primary cooling; max facility inlet water temperature 45 °C |
| Rack Power Draw (TDP) | ~120 kW total rack power (reference configuration) |
| Power Distribution | Rack-integrated high-voltage DC (HVDC) or AC power shelf architecture per MGX spec |
| Form Factor | Full-rack (42U–56U depending on OEM build); NVIDIA MGX standardized mechanical envelope |
| Management | BMC-based out-of-band management per node; NVIDIA Base Command Manager; DCGM GPU telemetry |
| Software Ecosystem | CUDA, cuDNN, TensorRT, NEMO Framework, NVIDIA AI Enterprise, Triton Inference Server |
| Launch | Q1 2025 |
The NVIDIA MGX GB200 NVL72 Reference Server Platform is built for enterprise data-center workloads — virtualization (VMware, Proxmox, Nutanix), private cloud, database hosting, and AI/ML training. It fits standard EIA-310 server racks and supports redundant PSUs and hot-swap drives common in production environments.
Key specifications for the NVIDIA MGX GB200 NVL72 Reference Server Platform: new condition; manufacturer NVIDIA; platform architecture NVIDIA MGX GB200 NVL72 — rack-scale reference design; gpu count per rack 72× NVIDIA Blackwell B200 GPUs; cpu count per rack 36× NVIDIA Grace CPUs (72 Arm Neoverse V2 cores each); superchip configuration 36× Grace Blackwell Superchips (1× Grace + 2× B200 per Superchip); ai compute (fp4) 1.4 ExaFLOPS. Manufacturer part number MGX-GB200-NVL72. For the full datasheet with electrical, environmental, and compliance details, contact our pre-sales engineering team.