NVIDIA GB200 NVL72 GPU Infrastructure & Deployment_
GB200 NVL72 combines 72 Blackwell GPUs and 36 Grace CPUs in a rack-scale NVLink domain. Liquid-cooled compute and high-density cabling require an approved OEM/site installation plan.
What Is the NVIDIA GB200?_
The NVIDIA GB200 NVL72 is a Blackwell rack-scale system with 72 GPUs, 36 Grace CPUs and fifth-generation NVLink. NVIDIA lists 13.4 TB of aggregate GPU memory. Its DGX reference rack uses 18 compute trays, each with four GPUs and two CPUs; exact OEM configurations must be checked against the supplied documentation.
NVIDIA describes roughly 120 kW consumption for the DGX GB reference rack. CPUs and GPUs use liquid cold plates, while other components still depend on airflow. Qualified facility/OEM teams must approve power, cooling and placement requirements; Leviathan can scope physical assembly, cabling and handoff without offering any liquid-side work.
Technical Specifications_
| Specification | GB200 NVL72 |
|---|---|
| Architecture | Blackwell |
| GPU Memory | 13.4 TB HBM3e per NVL72 rack (NVIDIA published specification) |
| GPUs per Rack | 72 (NVL72 configuration) |
| TDP per Rack | Approximately 120 kW in NVIDIA DGX GB reference; verify OEM input |
| Rack Weight | Exact OEM populated/shipping weight required |
| Interconnect | NVLink 5.0, InfiniBand NDR |
| Networking | 400G ConnectX-7 in DGX GB200 reference; OEM BOM governs |
| Cooling | Liquid-cooled compute; ancillary airflow still required |
Deployment Considerations_
Use the exact OEM electrical specification and approved facility design for rack input, upstream capacity and protection. Hardwiring, busway taps, new circuits and energization are outside Leviathan scope. Existing-equipment/rack-PDU cord work requires separate OEM/site/personnel/insurance agreement.
The DGX GB200 reference uses ConnectX-7 400G adapters for scale-out networking. Agree the actual switch/adapter BOM, compatible connector and cable reach, polarity, port map and pathways before installation; NVLink scale-up and external fabric cabling are distinct systems.
NVL72 compute requires its OEM liquid-cooling system and approved site interface. The customer-appointed mechanical/OEM specialists own CDU work, manifolds, fluid connections, leak/pressure tests and loop commissioning. Leviathan provides no liquid-side service.
Release the GB200 connection at the right physical boundary
A GB200 NVL72 rack does not have one interchangeable kind of fabric connection. In the named NVIDIA DGX GB system, the rack-internal NVLink path uses tray connections and a passive copper cartridge backplane. The compute tray also has external cluster, storage and management interfaces. An external fabric scope must distinguish those boundaries before a cable description can become a physical work item.
NVIDIA's September 17, 2026 hardware guide identifies four single-port ConnectX-7 cluster NICs per DGX GB200 compute tray, separate BlueField-3 storage and management hardware, and BMC interfaces. The companion networking guide distinguishes the OSFP cluster port from QSFP BlueField ports and RJ45 management ports. These are named DGX source descriptions, not every OEM GB200 configuration, a supported cable list or proof that the delivered rack needs new work at all those interfaces.
For release, attach the exact delivered tray and port identity to the intended network role, remote endpoint, selected supported assembly or channel, permitted path and work boundary. Identify any supplier-completed rack connections separately from field connections. A GPU population, connector shape or network name alone supplies none of those missing associations. The network owner must identify the intended operating configuration; this page does not configure it.
| Named boundary | What the source identifies | What a release still needs |
|---|---|---|
| Rack-internal NVLink | Tray connections through the rack's copper cable-cartridge backplane. | Actual supplied assembly, OEM task boundary and any permitted field work; not a substitute for external fabric links. |
| Cluster scale-out | ConnectX-7 OSFP ports assigned to computenet in the DGX GB200 networking table. | Actual tray/port, remote switch/port, supported channel and released path. |
| Storage and in-band management | The source distinguishes BlueField-3 left and right QSFP port roles. | Exact card and port orientation from the applicable view, intended role and remote endpoint; a connector name does not decide these. |
| Out-of-band management | Named BMC RJ45 interfaces; the same table marks the separate LAN port Not Connected. | The actual BMC endpoint and customer-released management connection, rather than an assumption that every RJ45 is the intended port. |
Source facts: NVIDIA DGX GB hardware and model-specific networking tables, checked October 19, 2026. No port configuration, firmware choice, cabling method, topology approval, supported-part selection or actual site permission is supplied.
Resolve a missing destination and a wrong port before release
Fictional EX-GB138 contains two intended field connection records for a DGX GB200 compute tray identified as CT07. Record C01 names Left Bay OSFP P3 and the cluster role, but its remote switch port and exact channel are absent. Record C02 calls the port marked LAN an out-of-band BMC connection. No live rack, observed port, cable order or customer project is represented.
The first record has an incomplete connection definition: a plausible named local cluster endpoint does not identify its other end or supported path. The second record has a specific source conflict: the cited compute-tray table marks LAN Not Connected and separately identifies the BMC management interface. That conflict does not authorize repatching or prove a fault. The actual equipment and network owners must resolve the intended endpoint against the delivered configuration.
| Record | Known input | Required decision and return |
|---|---|---|
| C01 | CT07, Left Bay OSFP P3, cluster role. Remote endpoint and channel not returned. | The network and equipment owners identify the peer and supported complete channel; the physical scope then identifies the permitted route and task. |
| C02 | CT07 LAN described as BMC management, conflicting with the named source table. | Resolve the actual management endpoint and applicable OEM view. Retain the superseded description with its correction instead of silently treating a different port as observed. |
The return should identify the governing source and connection revision, exact local and remote endpoints, assembly or channel, supplied-versus-field boundary, permitted route, responsible decision owner and included check or handoff. Keep unresolved items visible. Two intended records are not two installed cables or passing links. Physical observations, instrument results if separately agreed, network operation and customer acceptance remain distinct evidence.
Related Resources
Discuss Your Infrastructure Project_
Tell us about your project, and our team will follow up to discuss the scope, requirements and availability.
Discuss a ProjectEmail Project Scope