Comparisons
B200 vs GB200: GPU, HGX, DGX and NVL72 Names Explained
B200 identifies a GPU; HGX B200 identifies an eight-GPU platform, and DGX B200 identifies a complete NVIDIA server built with eight B200 GPUs. A GB200 Grace Blackwell Superchip combines one Grace CPU with two Blackwell GPUs. GB200 NVL72 is a rack-scale system with 36 Grace CPUs and 72 Blackwell GPUs. These names describe different units. Establish the complete delivered system before comparing quantities, site requirements or installation responsibilities.
Identify the unit behind the name
A proposal headed B200 or GB200 is incomplete as an installation description. The missing word may be GPU, baseboard, server, Superchip, compute tray or rack. Each changes what the quantity counts and which product documents apply. Ask for the full manufacturer model and delivered configuration before carrying the description into an equipment schedule.
Use the following distinctions to decode the name. The hardware counts refer to the cited NVIDIA products. They do not define every OEM implementation, order bundle or permitted field task.
Scroll the table horizontally on smaller screens.
| Name | Product level and documented identity | What still needs identifying | Source |
|---|---|---|---|
| B200 GPU | One Blackwell accelerator; not a complete CPU host or server | The supported system, installed GPU quantity and full configuration | [N5] |
| HGX B200 | An eight-B200-GPU baseboard/platform used in complete systems | The OEM server model, CPU, options and delivered integration state | [N5] |
| DGX B200 | A complete NVIDIA system with eight B200 GPUs and two Intel Xeon CPUs | The ordered options, supporting infrastructure and project scope | [N3] |
| GB200 Grace Blackwell Superchip | One Grace CPU connected to two Blackwell GPUs | The complete system containing the Superchips | [N1] |
| DGX GB200 compute tray | A tray with two Grace CPUs and four Blackwell GPUs | Its place in the specified rack and its authorized service boundary | [N2] |
| GB200 NVL72 | A rack-scale system with 36 Grace CPUs and 72 Blackwell GPUs | The exact system vendor, configuration, delivery state and facility dependencies | [N1] |
A platform name explains an architecture level. The purchase configuration and applicable manufacturer documents define the actual delivered equipment.
Read HGX and DGX as part of the complete model
HGX B200 is the platform inside a system, so an HGX B200 description does not choose one chassis size or server configuration. NVIDIA's reference architecture identifies the eight-GPU baseboard and describes configurable OEM systems. The current HGX specifications also identify fifth-generation NVLink for HGX B200; it should not be described as an NVLink 4 platform.
Two complete-system documents show why the missing model matters. Lenovo lists its ThinkSystem SR680a V3 with B200 as an 8U server. NVIDIA lists DGX B200 as a 10U rackmount system. Both contain eight B200 GPUs, but the shared accelerator count does not make their mechanical specifications the same. This is a model-specific example, not a rack-fit determination or an exhaustive list of B200 systems.
DGX also needs its suffix. DGX B200 is the eight-GPU server described in its own user guide; DGX GB200 refers to a different rack-scale system. A request for one DGX without the rest of the name leaves the delivered unit unresolved. Use the full model on the equipment list, installation brief and document register. The separate HGX-versus-DGX guide covers broader integration and support questions.
Separate the GB200 Superchip from the tray and rack
The Superchip is a CPU-and-GPU building block, not the whole rack. In NVIDIA's DGX GB200 implementation, one compute tray contains two Grace CPUs and four Blackwell GPUs. That is the CPU/GPU count of two GB200 Superchips. The rack documentation lists eighteen compute trays. Calling each tray one Superchip would undercount the documented compute hardware by a factor of two.
The arithmetic below is useful for reconciling a product description with an inventory. It does not establish the number of shipping containers, loose components, customer-replaceable units or field assembly steps. Those require the actual shipment and service documentation. Do not turn a count reconciliation into an instruction to disassemble, install or replace hardware.
Scroll the table horizontally on smaller screens.
| Named unit | Count per named unit | Reconciliation to the rack | Source |
|---|---|---|---|
| GB200 Superchip | 1 Grace CPU + 2 Blackwell GPUs | 36 Superchips: 36 CPUs and 72 GPUs | [N1] |
| DGX GB200 compute tray | 2 Grace CPUs + 4 Blackwell GPUs | 18 trays: 36 CPUs and 72 GPUs | [N2] |
| GB200 NVL72 rack-scale system | 36 Grace CPUs + 72 Blackwell GPUs | One 72-GPU NVLink domain | [N1] |
The tray count describes the cited NVIDIA DGX implementation. It is not a universal shipping, packaging or field-service specification for every vendor.
Keep equal GPU counts separate from system equivalence
Consider a hypothetical equipment list for 72 GPUs. If it specifies complete eight-GPU DGX B200 servers, the arithmetic is nine servers. A documented GB200 NVL72 rack also contains 72 GPUs. The common total does not turn those nine servers into one NVL72 domain, and it does not make the Blackwell GPUs inside GB200 interchangeable field parts with B200 modules in an HGX system.
The CPU composition is also different: DGX B200 contains Intel Xeon CPUs, while GB200 Superchips incorporate Grace CPUs. The platform owner must review the relevant software and application requirements for the actual system. A shared GPU architecture name is insufficient evidence of a supported software configuration or a migration path.
Separate the selection questions. Memory capacity belongs to an exact GPU/system specification. Workload speed belongs to a reproducible benchmark with its workload, precision, software and system count stated. Deployment readiness belongs to the complete system and site package. A marketing name or equal accelerator total cannot answer all three. This guide establishes identity; it does not rank performance, select workload capacity or promise a drop-in conversion.
Resolve ambiguous equipment lines before the delivery handoff
Retain the supplier's exact description, then add a plain-language unit field. Do not silently replace GB200 with rack or HGX B200 with server. If the unit is unclear, leave it unresolved until the supplying party identifies it. This prevents a rough project description from becoming an unsupported equipment or labor quantity.
The table below is an original review aid for names that need clarification. It asks for evidence; it does not prescribe a purchasing process or change anyone's contract. Once the line is resolved, carry the same identity into the receiving record and physical work package. Keep private order references and site drawings in the project's controlled records.
Scroll the table horizontally on smaller screens.
| Ambiguous description | Clarification to request | Record to retain | Source |
|---|---|---|---|
| 8 x B200 | Eight GPUs inside which complete system, or a different unit? | Full system model, configured GPU count and supplier line-item explanation | Planning analysis |
| 1 x HGX B200 | Baseboard/platform, configured OEM server or a larger bundle? | Exact manufacturer model and configuration revision | Planning analysis |
| 1 x DGX | DGX B200 server, DGX GB200 rack-scale system or another model? | Complete product name and ordered configuration | Planning analysis |
| 18 x GB200 | Superchips, compute trays or another supplier-defined unit? | Named unit, per-unit CPU/GPU count and complete-system relationship | Planning analysis |
| 1 x GB200 NVL72 | Which vendor implementation and delivery/integration state? | System identity, included equipment list and responsible parties for remaining work | Planning analysis |
The examples are hypothetical descriptions, not actual Leviathan orders or customer records.
Assign the work after the system is known
A resolved hardware name is the starting point for a scope discussion, not authorization for every activity around it. Ask the system supplier or designated integrator to identify what arrives completed, what remains to be done and which work is restricted to an authorized service party. Confirm the document revision that governs the ordered configuration. Avoid blanket assumptions that every rack is factory-complete, every internal component is field-serviceable, or every site team may perform the same work.
Facility electrical and cooling readiness need their own responsible owners and applicable system requirements. Physical placement, mounting and cabling need the released site plan and agreed work boundary. System bring-up, software qualification and workload acceptance need separately identified owners and criteria. The product name does not assign those responsibilities.
For a Leviathan scope discussion, provide the complete equipment model, quantities, site location, proposed schedule and available approved drawings or connection records. Leviathan can discuss agreed rack-and-stack and physical cabling work with customer-furnished equipment, subject to project requirements and availability. Identify the appointed equipment/material supplier and delivery responsibilities; this comparison is not a quotation or stock-availability statement. Electrical installation and liquid cooling installation remain excluded, and system commissioning is not a standard service. The linked readiness and RFP guides develop the physical work package once the equipment identity is settled.
References: [L1]
Common planning questions
Is GB200 simply another name for a B200 GPU?
No. The GB200 Grace Blackwell Superchip combines a Grace CPU and two Blackwell GPUs. A B200 GPU, an HGX B200 platform and a GB200 NVL72 rack describe different product levels. Use the complete system and configuration when assessing compatibility or deployment.
Does one HGX B200 mean one complete server?
The HGX B200 name identifies an eight-GPU baseboard/platform. A supplier may describe a complete OEM system using that platform, but the full server model and configuration still need to be stated. Do not infer a chassis size or included accessories from HGX alone.
How many GB200 Superchips are in an NVL72 configuration?
The published CPU/GPU configuration corresponds to 36 Superchips: one Grace CPU and two Blackwell GPUs each. In the cited DGX GB200 implementation, eighteen compute trays each contain two Grace CPUs and four GPUs. A tray is not one Superchip.
Does this comparison authorize a GB200 installation procedure?
No. It explains equipment identity and the evidence needed for a scope discussion. Actual site requirements, service permissions, handling, electrical and cooling work, system checks and acceptance remain with the applicable qualified owners and approved project scope.
Primary sources
- [N1] NVIDIA GB200 NVL72: overview and configuration specificationsChecked 2026-10-15. Adopts Superchip and NVL72 CPU/GPU identity and domain only. No projected performance, availability, facility design or service procedure adopted.
- [N2] NVIDIA DGX GB Rack Scale Systems User Guide: hardwareChecked 2026-10-15. Uses DGX GB200 compute-tray and rack counts only. The guide also covers GB300; GB300 comparison, operational commands, power and liquid procedures are outside this article.
- [N3] NVIDIA DGX B200 User Guide: introduction and hardware overviewChecked 2026-10-15. Eight B200 GPUs, two Intel Xeon CPUs and 10U mechanical identity apply to this named NVIDIA system. Not a generic HGX chassis, rack-fit or installation approval.
- [N4] NVIDIA HGX platform: HGX B200 specification columnChecked 2026-10-15. Uses the HGX B200 column for eight Blackwell SXMs and fifth-generation NVLink. Other generations and marketing performance claims are not compared here.
- [N5] NVIDIA HGX AI Factory H100/H200/B200 reference architecture: componentsChecked 2026-10-15. Establishes the B200 eight-GPU baseboard/platform and OEM system configuration boundary. No reference-cluster sizing, performance or TCO claim adopted.
- [O1] Lenovo ThinkSystem SR680a V3 with B200 product guide, LP2247Checked 2026-10-15. Uses the named eight-B200, 8U server as a configuration-specific chassis example; the public guide was updated September 29, 2026. No inventory, support eligibility or site-fit claim.
- [L1] Leviathan Systems rack-and-stack service scopeChecked 2026-10-15. Current public service description, read alongside founder-approved operating boundaries. Not evidence of OEM authorization or excluded service capabilities.