Platforms
NVIDIA H100 vs H200 vs B200: Compare the Deployment Configuration
H100, H200 and B200 identify GPU products, not complete installation plans. Compare the exact variant and full server before carrying power, cooling, rack or cable assumptions into a deployment. In NVIDIA's cited eight-GPU HGX reference architecture, H100/H200 use fourth-generation NVLink and B200 uses fifth-generation NVLink. B200 is not universally liquid-cooled: NVIDIA documents an air-cooled DGX B200 system. The offered OEM configuration and approved site interfaces control the actual work.
Identify the product level before comparing deployment requirements
Begin with the thing being delivered. A GPU or SXM module is a component. An HGX baseboard combines GPUs and their associated interconnect in a platform used by system manufacturers. A DGX product is a named complete NVIDIA system. An OEM's HGX-based server adds its own chassis, host CPUs, networking, power supplies, cooling arrangement, service access and supported configuration. A rack can contain that server plus other equipment with separate requirements.
The distinction changes the questions a buyer can answer. A GPU specification can help establish a memory or interconnect requirement. It cannot identify the complete server's rail kit or authorize a position in an existing cabinet. A generic HGX reference drawing is also insufficient to confirm that the delivered OEM chassis has the same physical interfaces as another system using the same GPUs.
Write the full manufacturer and system model, GPU variant and count, configuration identifier and applicable document revision at the top of the comparison. Keep any rack-level equipment separate. If the request says only eight H200s or one B200 rack, return it for a more precise equipment identity before treating the work as repeatable.
Keep H100 and H200 variants visible
NVIDIA's current H100 product table distinguishes H100 SXM from the PCIe-form-factor H100 NVL. It lists 80 GB per GPU and 900 GB/s NVLink for SXM, versus 94 GB and 600 GB/s NVLink for NVL. Those values belong to the named columns. A request containing only H100 PCIe or H100 does not identify all the information that the installation team needs.
The H200 product page likewise separates SXM and NVL. It lists 141 GB HBM3e for both, while describing different form factors and server configurations; H200 NVL is identified as a PCIe, dual-slot air-cooled offering. The page still marks its specification columns preliminary and subject to change. Retain that qualification and obtain the current offered system specification rather than promoting a general web table into a contractual configuration.
These examples do not establish that an H200 module can replace an H100 module in a particular existing server. Similar product-family language, memory capacity or a shared nominal power rating is not an OEM-approved upgrade path. Ask the system supplier to identify the supported replacement configuration, required changes and responsible service organization in writing. Until that answer exists, describe a proposed refresh as a configuration review, not an approved drop-in swap.
Compare like-for-like HGX specifications
For a controlled comparison, NVIDIA's HGX H100/H200/B200 reference architecture provides a useful set of SXM configurations. It identifies the GPU-local memory separately from the eight-GPU node total and distinguishes the NVLink generation from the NVSwitch generation. The table below retains that exact reference scope. It does not summarize every H100, H200, Blackwell or Grace Blackwell product.
Memory capacity is a hardware-selection input. The additional capacity in one configuration does not, by itself, prescribe a new external cable plant or demonstrate that a site's cooling can support the complete server. Keep the workload fit decision with the application/platform owner, and use the complete system documentation to assess the physical interfaces. For broader capacity comparisons, use the separate GPU memory guide rather than mixing server and rack totals into these per-GPU rows.
References: [N3]
Scroll the table horizontally on smaller screens.
| Cited GPU variant | Memory per GPU | NVLink generation | Cited GPU-to-GPU bandwidth | Source |
|---|---|---|---|---|
| H100 SXM | 80 GB HBM3 | Fourth generation | 900 GB/s | [N3] |
| H200 SXM | 141 GB HBM3e | Fourth generation | 900 GB/s | [N3] |
| B200 SXM | 180 GB HBM3e | Fifth generation | 1,800 GB/s | [N3] |
Source N3, Components, Tables 1–2, checked October 15, 2026; source page last updated August 31, 2026. Values retain the source's units and product scope. They are published specifications, not workload benchmarks or available free memory.
Separate NVLink from the external network cable schedule
In the cited eight-GPU HGX configurations, H100/H200 combine fourth-generation NVLink with third-generation NVSwitch; B200 combines fifth-generation NVLink with fourth-generation NVSwitch. Mixing those two generation numbers produces an incorrect platform description. Keep the GPU interconnect, the switch component and the external network interfaces as separate fields.
The DGX B200 user guide illustrates the next boundary. Its cluster-network description identifies four OSFP ports serving eight ConnectX-7 single-port cards. That is a specific system port layout, not a rule that every B200 GPU has a set number of external fiber connectors. The guide also directs readers to the applicable adapter/firmware compatibility information for supported network cables and adapters.
Build the external cable schedule from the actual server/NIC and switch endpoints, selected interfaces, qualified media, intended paths and agreed spare policy. The higher NVLink figure in a component table does not select OS2 instead of multimode, establish reach or create a trunk quantity. Conversely, a complete external cabling handoff is not proof that the system's internal GPU interconnect has passed platform acceptance. Those are different evidence records with different owners.
Use complete-system documents for power and cooling
B200 does not imply a universal liquid-cooling requirement. NVIDIA explicitly describes air-cooled DGX B200 configurations on its current product page, and its user guide documents that complete system's environmental and airflow requirements. This establishes an air-cooled product example; it does not approve that product for every air-cooled facility or describe all OEM B200 designs. The proposed chassis and facility still need a configuration-specific review.
Complete-system numbers also answer a different question from GPU component ratings. The current DGX H100/H200 user guide lists an 8U system and 10.2 kW maximum in its power table; the DGX B200 guide lists 10U and 14.3 kW maximum. These examples show why a plan based on a GPU's nominal wattage can miss the actual delivered enclosure and whole-system requirement. They are not measured operating loads, circuit-sizing instructions or generic rack limits.
Request the actual OEM input, redundancy, environmental and cooling documents for the offered system, together with the facility owner's applicable interface approval. Carry the OEM-qualified accessory requirements into that review. Neither a common connector name nor an apparently similar power supply establishes that cords or other parts may be exchanged across products. Electrical design, connection authorization and any liquid-side work require their own qualified owners and approved scope.
Scroll the table horizontally on smaller screens.
| Named system documentation | Enclosure | Published maximum system power | How to use the comparison | Source |
|---|---|---|---|---|
| DGX H100/H200 user guide [N4] | 8U rackmount | 10.2 kW maximum | Confirm the actual ordered configuration and facility interfaces | [N4] |
| DGX B200 user guide [N5] | 10U rackmount | 14.3 kW maximum | Review the changed whole-system and enclosure requirements | [N5] |
Checked October 15, 2026. Source-specific examples only; no upstream electrical, redundancy or thermal design is performed by this table.
Compare the rack package, access and documentary interfaces
A rack elevation needs the actual enclosures, accessories and reserved access requirements. Do not fill a cabinet by dividing its nominal height by a marketing server height alone. The proposed arrangement still needs the approved equipment list, rack/accessory compatibility, service access and facility conditions. A deployment of the same GPU count can use a different chassis or accessory package and therefore require a different physical plan.
Keep shipment, receiving and installed configurations distinct. Identify whether equipment arrives in a complete rack, as complete servers or as separately controlled packages. Ask the responsible parties for the actual weights, dimensions and approved handling plan; this guide supplies no generic floor rating, rack mass, lift method or allowance for moving loaded cabinets.
The useful comparison is a documented change list: which enclosure, accessory, cable endpoint, access condition or handoff requirement changes between the existing installation and the proposed one? Keep unresolved items beside the affected system. This allows a qualified reviewer to close the particular gap without treating every difference as a demand to rebuild the entire facility.
Request a release package that can be reviewed
Use a document request before agreeing an installation window. The worksheet below is an original coordination aid: each requested record should identify the actual system and the party responsible for its acceptance. It is not a substitute for the manufacturer's installation instructions or an approved design.
For a refresh, keep the current and proposed configurations visible in the same review. If an accessory or interface is intended to remain, ask the responsible owner to identify the evidence that supports reuse. If the delivered configuration changes, reopen the affected interface rather than carrying an old release into the new workfront. Record both the remaining question and the person who can close it.
Scroll the table horizontally on smaller screens.
| Record to request | Configuration-specific question | Owner to identify | Source |
|---|---|---|---|
| Delivered system identity | Which manufacturer, full model, GPU variant/count and configuration are being delivered? | Customer equipment owner and supplier | Planning analysis |
| Current OEM document set | Which revisions cover this enclosure, supported accessories and service conditions? | System supplier/OEM | Planning analysis |
| Rack and access package | Which rack positions, accessories, interfaces and access clearances are approved? | Customer/facility and qualified design owners | Planning analysis |
| Facility-interface approval | Which system-specific power, cooling and environmental requirements have been accepted? | Facility and appointed qualified technical owners | Planning analysis |
| External endpoint schedule | Which actual NIC/switch ports, media, paths and labels govern the physical cable work? | Network/design and installation-package owners | Planning analysis |
| Work authorization and materials | Which tasks, workfront, customer-furnished items and access permissions are released? | Customer project and facility owners | Planning analysis |
| Separate acceptance records | Who accepts physical installation, platform bring-up, software and workload performance? | Named customer acceptance owners | Planning analysis |
Open items stay open until their responsible owner provides applicable evidence. A completed worksheet is not professional certification or permission to service equipment.
Work through a refresh request without assuming a swap
Consider a hypothetical request: replace an eight-GPU H100 installation with H200 and keep the rack layout and cable package unchanged. The first question is whether the proposal concerns component replacement in an existing supported server or delivery of a different complete system. The two requests create different approval and execution paths, even when the final GPU count is eight.
The equipment owner obtains the supplier's exact supported configuration. The facility reviewer evaluates the actual offered system interfaces. The network owner confirms whether the same external endpoint schedule applies, and the physical project owner checks the released layout and materials. Any reuse decision stays tied to those records; it is not justified by the sentence that both GPUs are Hopper.
For a change to B200, follow the same evidence sequence with the actual offered system. The DGX examples above show that whole-system specifications can change, but do not predict the differences for another manufacturer's enclosure. The outcome should be a precise list of approved reuse, required changes and open conditions, rather than a universal claim that the upgrade is either effortless or impossible. This is a planning example, not a customer case study.
Separate physical handoff from platform acceptance
A physical installation record can identify installed equipment, completed cable/label work, agreed inspection observations, documented defects and the released drawing set. Platform acceptance addresses different questions: supported software and firmware, component behavior, internal interconnect validation, external fabric operation and workload outcomes. Agree the accepting team for each record before the installation window.
Do not use a visible power indicator or one detected GPU count as a substitute for the customer's complete platform acceptance. Likewise, a workload result does not automatically close missing physical labels or handoff documents. Keep the records connected by the actual equipment identifiers while retaining their separate scope and acceptance criteria.
Leviathan can discuss physical mounting, rack-and-stack installation, structured cabling, routing, dressing, labeling, inspection/remediation and agreed physical handoff documentation. The actual site, customer-furnished equipment, authorized work and available resources determine the engagement. Identify the equipment/material supplier, approved configuration, quantities and delivery responsibilities before agreeing the physical work. Electrical installation and liquid-side work remain excluded; full software commissioning is not a standard service, and workload tuning and benchmark acceptance belong with the appointed platform owners. Send the exact system and installation package so a useful scope discussion can begin.
References: [L1]
Common planning questions
Can H200 replace H100 without changing the installation?
A family comparison does not approve a replacement. Obtain the actual OEM-supported configuration and review the complete system, retained interfaces, software/service requirements and facility approval before treating any part of the plan as reusable.
Does every B200 system require liquid cooling?
No. NVIDIA documents an air-cooled DGX B200 configuration. That fact does not establish compatibility with every air-cooled site or describe every OEM B200 server; use the actual system and facility documents.
Which NVLink generations do these HGX configurations use?
The cited HGX H100/H200 reference configurations use fourth-generation NVLink; HGX B200 uses fifth-generation NVLink. Keep these labels separate from NVSwitch component generations and external networking interfaces.
Does B200's NVLink bandwidth determine the external fiber type?
No. Specify external cabling from the actual server/NIC and switch endpoints, supported interfaces/media, paths and released design. An internal interconnect figure does not establish reach, fiber mode or trunk count.
Primary sources
- [N1] NVIDIA H100 product specifications: SXM and NVLChecked 2026-10-15. Current named SXM/NVL columns only; 80/94GB memory, 900/600GB/s NVLink and form factors. No all-H100 variant summary, speedup, pricing or compatibility promise adopted.
- [N2] NVIDIA H200 product specifications: SXM and NVLChecked 2026-10-15. Named SXM/NVL product distinctions and 141 GB HBM3e. Page footnote 1 still says preliminary specifications, subject to change; retained in copy. No product replacement or facility compatibility inferred.
- [N3] NVIDIA HGX H100/H200/B200 reference architecture: ComponentsChecked 2026-10-15. Tables 1–2, exact SXM reference configurations and fourth/fifth NVLink versus third/fourth NVSwitch. Page last updated August 31, 2026. No generalization to all GPU-family or OEM-system variants.
- [N4] NVIDIA DGX H100/H200 user guide: system introduction and specificationsChecked 2026-10-15. Complete-system identity, 8U mechanical specification and 10.2 kW maximum system-power example. Not a universal rack requirement, measured load or electrical design instruction.
- [N5] NVIDIA DGX B200 user guide: system, network and interface specificationsChecked 2026-10-15. Eight-GPU complete system, 10U, 14.3 kW maximum, exact OSFP/ConnectX-7 mapping and supported-accessory references. Page last updated September 8, 2026. No servicing, live electrical, handling or liquid procedure reproduced.
- [N6] NVIDIA DGX B200 product pageChecked 2026-10-15. Current page explicitly describes air-cooled DGX B200 configurations in its performance figure captions. Only this system configuration fact is adopted; projected performance, pricing and TCO claims are excluded.
- [L1] Leviathan rack-and-stack service scopeChecked 2026-10-15. Published bounded physical-service description and founder-approved operational boundaries. Not evidence of particular staff, qualifications, tools, site authorization or guaranteed availability.