CARDINALA GPU fits a server only when the complete configuration is supported. Physical space is one check. Power delivery, cooling, firmware and software are separate checks. Build a configuration record before buying, then ask the server manufacturer to confirm it.
Start with the server's exact identity
A server family name is a starting point, not a complete compatibility statement. Record the machine type or service identifier and the installed configuration. Include CPU count, riser options, power supplies and cooling components. Two machines with the same front-panel name can have different expansion options.
Then identify the proposed accelerator by OEM part number, memory configuration and physical form. Keep that record separate from the marketing family. An H100 PCIe 80GB and an H100 SXM5 80GB require different installation paths.
OEM product guides illustrate why this detail matters. Lenovo's H100 guide separates ordering information, server support and operating-system support. Use that structure to form the questions for the actual server you intend to buy or modify. Source: Lenovo H100 PCIe product guide
Ask for a written configuration confirmation that names both the server and accelerator parts. A reply saying only that the server “supports H100” leaves the variant and required options unresolved. Preserve the answer with the commercial offer.
Scroll across the diagram to read every label
Download diagramPhysical fit includes more than card length
For PCIe cards, check height, length and occupied slot width. Include the bracket and any extender. Leave room for the power cable and interconnect hardware. A bare-board dimension may not describe the clearance needed by the installed assembly.
Record the required riser and its slot population rules. Ask whether adjacent devices affect the supported arrangement. A card may fit in an empty slot while the complete populated configuration remains unsupported.
Distinguish a physical x16 slot from its electrical connection. Confirm the active lane width and supported generation for the proposed slot. Also identify which CPU or PCIe switch serves it. This information can matter for both compatibility and the application's communication paths.
For a pair of bridged cards, check the exact bridge and spacing. NVIDIA's H100 PCIe brief describes bridge placement and additional clearance requirements. Do not assume a bridge from another card or slot arrangement is suitable. Source: H100 PCIe mechanical and bridge sections
Use the installed assembly as the unit of review
Ask the seller to show how the card is retained in the chassis. Identify brackets, support pieces and cable routing in the component list. Missing small parts can turn an apparently complete accelerator offer into an unfinished installation.
For an existing server, compare the proposed arrangement with the actual machine. Do not rely on a generic photograph from the server's product page. If a riser or cooling kit must change, price and document that change before comparing the offer with another system.
For a used complete server, request interior photographs and the installed-component inventory. Those records help establish what is present. They do not replace the OEM's supported-configuration information.
Confirm the supported power path
A sufficiently large power-supply label does not establish that every card can receive the required power. The cable, connector, riser and system configuration also matter. Ask the OEM which approved parts are required for the proposed accelerator count.
Dell's R760xa GPU-kit documentation includes specific power-cable and cooling guidance and cautions against unvalidated devices. This is an example of configuration-level requirements, not a blanket claim about other Dell servers. Source: Dell R760xa GPU kit
Record any supported power limit as part of the configuration. If a proposal only works with a lower limit, that condition belongs beside its performance evidence. Do not compare it with an unrestricted benchmark without explaining the difference.
Keep the complete server's requirements separate from the sum of accelerator ratings. Ask the facility operator to review the supported server configuration and installation plan. Use the power and cooling guide for that assessment rather than treating this compatibility checklist as a facility design.
Cooling is a qualification condition
Passive accelerator cooling relies on the server's airflow. Confirm the correct fans, air guides and supported inlet conditions through the OEM documentation. The absence of a fan on a card does not mean that it can operate without directed cooling.
Cooling options can also change the physical configuration. NVIDIA lists both air-cooled and liquid-cooled forms for A100 PCIe 80GB. Ask for the actual part number and cooling arrangement. Do not treat these as interchangeable versions of the same installation. Source: A100 form-factor specifications
If a seller removed accelerators from a different system, establish which thermal and mounting components remain with them. An OEM cooling assembly may belong to the source server's design. Confirm the required target assembly instead of improvising from whatever is included.
A detected device is not a fully validated system
Firmware and software must support the proposed hardware arrangement. Record BIOS or UEFI, management firmware, device firmware and driver versions. Ask which combination the OEM or platform supplier qualifies.
Do not apply a generic list of firmware settings to every server. Features affecting address mapping, virtualization or peer access depend on the platform and intended use. Use the documented configuration for that system, with a record of any changes.
The operating model matters. A bare-metal workload and a virtual-machine passthrough workload may have different requirements. Ask the team responsible for the deployment to confirm the intended mode before treating a seller's demonstration as relevant evidence.
NVIDIA's multi-GPU documentation describes topology and platform conditions for peer access. A system can enumerate its GPUs while still failing a communication requirement. Validate the capabilities the application actually needs. Source: CUDA multi-GPU platform requirements
Pin the application environment
Record the framework and application build, including any custom GPU extensions. Include the container digest when containers are used. A container helps preserve an environment, but it does not remove host-driver and hardware requirements.
CUDA's compatibility paths include limitations. Confirm the proposed driver and runtime combination against the official documentation, rather than relying on a general statement of CUDA support. Source: CUDA compatibility
For AMD or Intel platforms, use the relevant release-specific support information instead. Do not translate an NVIDIA qualification statement into an AMD or Gaudi approval. The AMD platform guide and Gaudi buying guide cover those software paths.
SXM and OAM are complete-platform decisions
A module is not a conventional PCIe add-in card. The H100 SXM5 reference and MI300X OAM reference identify different platform components. Their purchase requires the compatible baseboard and system around them.
Ask whether the quoted item includes only modules, a populated baseboard or a complete server. Keep these scopes separate in your comparison. A low module price cannot be compared directly with a server price without reconstructing the missing components.
Shared form-factor naming is not a field-upgrade approval. For example, MI325X and MI300X have different peak board ratings. Ask the OEM to confirm any proposed replacement using exact part numbers. The module category alone leaves power, firmware and cooling unresolved. Source: MI325X requirements
Intel's Gaudi readiness documentation similarly directs users to system-vendor power and thermal requirements. It distinguishes management access and accelerator-fabric connections in the described systems. That illustrates how much of a module deployment lives outside the accelerator itself. Source: Gaudi platform readiness
Choose the purchasing path that matches your starting point
You have a server with free expansion space
Start with its actual component inventory and supported options. Ask whether the target card is supported in that configuration. Identify every required change before requesting accelerator-only offers.
If the required changes are extensive, compare the retrofit with a complete qualified server. Include engineering time and the interruption to existing workloads in the decision. Do not assume that reusing a chassis automatically makes the project simpler.
You are replacing a failed accelerator
Use the service part number and approved replacement procedure. Ask whether the proposed replacement matches the supported configuration or requires another qualified part. A similar marketing name is not sufficient evidence.
Preserve the existing system's firmware and configuration record. Have the responsible operator follow the OEM service process. A purchase guide should not substitute for instructions governing work inside the actual server.
You are buying a complete used GPU server
Request a bill of materials and configuration export, not only the accelerator count. Confirm management access, host memory, storage and networking. Ask which components are included in the stated condition and any stated service arrangement.
Require an application demonstration that matches the proposed configuration. If the seller uses a different server or software image, record that limitation. The demonstration may still be informative, but it does not prove the delivered system's behavior.
Build a compatibility record before ordering
| Record | Evidence | Who resolves uncertainty |
|---|---|---|
| Server identity | Machine identifier and installed-component inventory | Seller and buyer's operator |
| Accelerator identity | Exact variant, OEM part number and revision | Seller, checked against manufacturer documentation |
| Physical installation | Riser, slot map and mounting components | Server OEM or qualified integrator |
| Electrical and thermal support | Approved cables, cooling configuration and power conditions | Server OEM and facility operator |
| Software configuration | Firmware, driver and application versions | Deployment owner |
| Acceptance evidence | Agreed checks tied to the delivered configuration | Buyer and seller |
Mark unresolved fields explicitly. An unknown part number or missing confirmation should not disappear when the proposal becomes a quote. Keep the technical record attached to the accepted offer so that substitutions can be reviewed against it.
Separate compatibility from condition. A perfectly matched part can still need condition evidence, while a seller's passing test can describe an unsupported target installation. Use the testing and acceptance guide for the second question.
Cardinal reviews seller-supplied documents and sources against the requested configuration. It does not claim to qualify server designs or test hardware. A precise compatibility record helps suppliers provide an offer that your technical team can evaluate.
Questions buyers ask about GPU server compatibility
Does an available PCIe x16 slot prove compatibility?
No. Confirm electrical width, space, power delivery, cooling and supported firmware. The OEM's configuration information is the appropriate starting point.
Can a larger power supply solve an unsupported configuration?
Not by itself. The supported power path and thermal arrangement still matter. Ask which complete configuration is approved for the proposed devices.
Can SXM or OAM modules be used through a generic PCIe adapter?
Do not assume a supported deployment from an adapter claim. Require the module's qualified platform and a documented integration path. The module is not a standard add-in card.
Does a successful boot prove the workload will run?
No. It establishes only part of the system behavior. Validate the required software, communication capabilities and application configuration separately.
What is the most useful information to send with a quote request?
Send the exact server configuration, target accelerator variant, quantity and destination. Add the software environment and required acceptance evidence. State which compatibility questions remain unresolved.