Field Guide · GB200 NVL72_
GB200 NVL72 Installation Checklist_
Before committing to installation, identify the delivered rack configuration, obtain the site readiness decisions, reconcile the scoped connections and agree what evidence completes the work. Those four decisions make a GB200 NVL72 package reviewable. A power headline or a total cable count does not.
Published · Updated
Use this checklist to scope and review physical work. It is not a rack assembly, energization or cooling procedure. The examples reference NVIDIA DGX documentation; another supplier's GB200 system needs its own current documents and authorized work instructions.
Start with the rack that will arrive
The cited DGX guide describes 18 compute trays, each containing two Grace CPUs and four Blackwell GPUs, and nine NVLink switch trays. It also describes an internal passive copper cable-cartridge backplane. These are properties of the documented rack architecture, not instructions to assemble every shipment from loose components. NVIDIA hardware reference
Put the supplier configuration and shipment contents beside the proposed installation scope. An integrated rack may arrive with work already completed; a different supply arrangement may allocate preparation elsewhere. Ask which assemblies are supplied installed, what receiving inspection is permitted, and who owns any discrepancy. Avoid charging for assumed tray fitting or opening integrated assemblies without authority.
Keep packed weight, installed weight and the authorized movement configuration distinct. Obtain the actual values and route constraints from the supplier and facility. A weight copied from another enclosure cannot release the delivery path or select handling equipment. NVIDIA's safety guidance requires technically qualified personnel and gives product-specific documentation precedence when documents conflict. Safety and applicability
Five records to resolve before mobilization
Record each item as received, under review, released for the named task, or open with an owner. Attach the revision and any conditions. “Document received” is not the same as approval to proceed. If only part of the work is released, identify that part explicitly rather than treating the entire rack as ready.
| Record | Required content | Responsible interface | Decision |
|---|---|---|---|
| Delivered equipment | Supplier, rack and tray identifiers, ordered configuration, shipment contents and integration state; current OEM installation and safety documents. | Customer and equipment supplier | What arrives assembled, what remains to be fitted, and who is authorized to do it? |
| Receipt and placement | Delivery route, packaged and installed dimensions/weights, access window, receiving condition record and approved handling/placement plan. | Customer, facility and qualified handling team | Can this shipment reach the agreed location using approved resources? |
| Facility interfaces | Exact rack power and thermal requirements; accepted electrical/mechanical designs; named readiness releases and any conditions. | Facility and customer-appointed electrical/mechanical specialists | Which dependencies are released, which are open, and which tasks may proceed? |
| Physical connections | Versioned endpoint schedule, network role, port identifiers, approved cable/optic part numbers, pathway, label scheme and permitted equipment state. | Customer network lead and installation scope owner | Is each scoped connection identified without substituting parts or guessing endpoints? |
| Checks and acceptance | Agreed inspection/test matrix, equipment and method, required records, acceptance owner and exception process. | Customer acceptance owner and responsible delivery teams | What evidence closes the physical package, and what belongs to a different acceptance stage? |
This is an original planning register, not an OEM acceptance form. For a broader facility survey, use the GPU site-survey checklist. Keep approvals and restricted site information in the project's authorized records.
Separate internal rack links from the field cable schedule
NVIDIA distinguishes the internal NVLink domain from compute communication across racks and from storage and management networks. A single combined cable count obscures those roles and the boundary between supplier integration and site installation. Network-role reference
For each field-installed connection, identify both physical endpoints, its network role, the approved assembly and relevant document revision. Keep installed configuration separate from available connector positions. An unused port is not automatically a missing cable, and matching connector shapes do not establish compatibility.
Keep internal backplane service work in its separately authorized OEM work package. For the external package, the customer's network design determines the topology, protocol and port use. The installation record should preserve those decisions without silently adding firmware, IP configuration or fabric acceptance to the cabling scope.
The DGX GB200 versus GB300 interface comparison addresses generation-specific port differences. This checklist uses that detail as an input to work readiness; it does not replace the comparison or create a second cable-selection guide.
Treat power and cooling as released interfaces
The shared DGX GB hardware guide describes approximately 120 kW rack consumption, liquid-cooled CPUs/GPUs and air-cooled networking/storage components. Those details explain why a simple “power available, liquid cooling available” checkbox is inadequate. They do not set the facility design for a purchased configuration. Hardware scope and thermal context
Request an explicit readiness disposition from each responsible electrical and mechanical team against the exact rack requirements. Identify any conditions that affect access, placement, cable routing or permitted equipment state. Do not substitute the sum of power-supply ratings for a signed capacity decision, or interpret the absence of a visible leak as completed cooling acceptance.
This guide provides no coolant handling, pressure test, bonding threshold, feed-disconnection test or energization sequence. Those decisions belong to the actual approved design and operating method. Leviathan's physical package must identify the interfaces without taking on excluded liquid-side or facility electrical installation.
Define the evidence before calling the rack accepted
“Test and commission” is too broad to price or sign off as one line. Separate the physical records from formal measurements, hardware QA and system validation. Agree who performs each activity and who accepts its evidence; a completed activity in one row does not close the others.
| Scope | Evidence to agree | What it does not prove |
|---|---|---|
| Receiving and mounting | Item and condition record; agreed placement/mounting checks; deviations linked to equipment identifiers. | A clean receiving record does not qualify the floor, lifting method or electrical installation. |
| Physical cable package | Reconciled endpoint and label records, agreed visual/connectivity checks and accessible exception list. | A cable record or link indication does not demonstrate application performance or complete fabric acceptance. |
| Formal optical or copper tests | Only the agreed method, equipment, personnel, limits, calibration/traceability requirements and report format for this engagement. | OTDR, loss testing, endface inspection and copper certification are distinct activities; a generic requirement to test every link does not select the appropriate method. |
| Hardware burn-in and QA | Equipment configuration, authorized method, conditions, duration, tools, personnel, results, exceptions and agreed acceptance criteria. | A proposed burn-in scope is not a guaranteed workload, universal soak duration or certification of the entire cluster. |
| Facility and platform acceptance | Separate records from the responsible electrical, mechanical, OEM, network and software teams, referenced where they affect the physical handoff. | The physical installation record does not replace facility commissioning, fabric configuration or workload validation. |
These are proposed evidence categories, not universal pass thresholds. Read the acceptance-evidence guide when reviewing a result package, and define any formal method against the actual equipment and engagement.
Example: the shipment and the scope disagree
Consider a fictional planning review. The purchasing record describes an integrated rack, while the draft installation scope includes fitting all compute trays. The endpoint schedule also lists management connections that the supplier says are already installed. Neither description is enough to authorize additional handling or declare the connections complete.
The useful correction is a revised responsibility record: the supplier confirms delivered integration state; the customer identifies the approved inspection boundary; and the installation scope names only the remaining authorized tasks. The cable schedule distinguishes pre-installed assemblies from field-installed connections, with a receiving check and an exception owner for each unresolved discrepancy.
The outcome is a clearer work package, not a claim of fewer labor hours or a successful deployment. If the records still conflict, keep the affected activity open. Separately released work can proceed only within its own approved boundary.
Make the physical handoff usable
Link each completion record to a rack or connection identifier and the approved scope revision. Include receiving discrepancies, the reconciled cable schedule, agreed check results, accepted changes and open exceptions. Name the acceptance owner and distinguish work completed from work accepted with conditions. Photographs should support a specific record and be shared only through permitted channels.
A useful exception identifies the affected item, the observed condition, the responsible party and the decision still needed. “Rack complete” is not a substitute for that list. Refer to the as-built handoff package for the broader record structure.
Leviathan can scope equipment receiving and preparation, rack assembly, hardware burn-in/QA, physical deployment, structured cabling, relocation and lifecycle support. Customer-furnished equipment remains a valid arrangement. Methods, resources, site permissions and acceptance are agreed for each project; no owned facility, stock position, OEM authorization or previous GB200 deployment is implied.
Discuss the receiving and rack-assembly package, rack-and-stack scope or structured cabling that matches the remaining work. Liquid-side work and facility electrical installation remain excluded. BIOS, firmware, IP configuration, OS imaging and complete software commissioning are not standard services.
GB200 installation planning FAQ
Is a GB200 NVL72 installation a 2,000-cable patching job?
Do not use that headline as an external cable order. The cited DGX guide describes internal NVLink cable cartridges as well as separate compute, storage and management networks. Count the actual field-installed connections from the approved endpoint schedule and distinguish supplier-integrated assemblies from site work.
Can approximately 120 kW be used to approve the rack location?
No. NVIDIA's shared DGX GB hardware guide gives that approximate rack-power figure as product context. The purchased system, current OEM requirements and facility-approved design must establish the actual electrical and thermal interfaces. An approximate figure cannot approve capacity, protection, redundancy or a connection method.
What receiving, rack-assembly and hardware-QA work can be scoped?
Equipment receiving/preparation, rack assembly and scoped hardware burn-in/QA are part of the confirmed portfolio. The supplied configuration, resources, method, conditions, duration and acceptance records must be agreed for the engagement. Customer-furnished equipment is also valid. This does not imply stock availability, an owned integration facility, OEM authorization or a prior GB200 installation.
Does this checklist authorize power, cooling or software commissioning?
No. It organizes inputs and responsibilities for a physical work package. Liquid-side work and facility electrical installation remain excluded. BIOS, firmware, IP configuration, OS imaging and complete software commissioning are not standard services. Actual installation, handling and test methods require the appropriate OEM, facility and qualified delivery authority.
Primary sources and applicability
The three official pages below were checked on October 17, 2026. They support the bounded product and safety statements, not Leviathan project evidence or approval of a particular installation. The checklists, responsibility categories and fictional example are original planning aids.
- NVIDIA DGX GB Rack Scale Systems User Guide — Hardware
September 17, 2026 guide: rack composition, cooling arrangements, internal cable backplane and approximate power context. Shared GB200/GB300 material must be applied to the named configuration.
- NVIDIA DGX GB Rack Scale Systems User Guide — Networking
Internal NVLink and distinct compute, storage and management roles; model-specific maps remain necessary for an actual endpoint schedule.
- NVIDIA DGX GB Rack Scale Systems User Guide — Safety
Qualified personnel, product-document precedence, handling and electrical precautions. This checklist does not reproduce or replace an OEM operating procedure.
Scope the physical work around your GB200 rack
Share the supplier configuration, site requirements, remaining installation tasks and required handoff. Our team will follow up to discuss scope, requirements and availability.
Discuss a ProjectEmail Project Scope