SENICO Electronics

Component sourcing and inventory guides

H100 SXM Versus PCIe: Define the Server Service Boundary

Define the H100 SXM, PCIe card and complete server service boundary before comparing cooling, replacement scope and procurement offers.

Quick Answer

An H100 procurement request should identify the physical assembly that will be installed and the assembly that can be serviced. SXM modules, PCIe accelerator cards and complete servers create different mechanical, cooling and support obligations. The H100 family name alone does not define an acceptable replacement. Start with the server manufacturer's approved configuration and service procedure, then describe the exact scope in the quotation request.

This is a platform selection guide, not a listing for a discrete semiconductor. NVIDIA's H100 information distinguishes SXM and PCIe form factors. That distinction should remain visible from the engineering requirement through receiving inspection; a shared GPU architecture does not establish mechanical interchangeability.

Key Specifications and Decision Points

NVIDIA identifies H100 SXM in its HGX platform options, while its H100 NVL product brief describes a passive, dual-slot PCIe card. These are useful anchors for a service discussion. They do not identify every field-replaceable assembly offered by a particular server vendor. Ask the vendor to connect the accelerator designation to the actual chassis, baseboard and cooling revision being supplied.

Use an assembly drawing to settle ambiguous language. Photographs can help receiving teams identify a configuration, but they cannot establish connector assignments, allowable installation forces or thermal requirements. Those decisions need the appropriate manufacturer documents.

  • Record the purchased object: a GPU module, a PCIe card, an accelerator baseboard, or a complete server. Include all brackets, retainers and cooling hardware expected in the delivered assembly.
  • Establish the service object independently: ask which assembly the maintenance organization is authorized and equipped to replace, and which diagnostic evidence it requires before authorizing that replacement.
  • Obtain a cooling responsibility statement covering the supplied heat sink or cold plate, host airflow or coolant requirements, and the party validating the assembled system.

Application and Acceptance Review

Consider a maintenance team buying spares for several existing AI servers. One machine may expose an approved PCIe card replacement procedure, while another may require exchanging a larger accelerator assembly. Ordering an apparently economical loose module without checking the procedure can leave the team with material it cannot qualify or install. Work backward from the service manual to the purchase specification.

Design the acceptance report around the service boundary. If a complete server is purchased, a loose-card diagnostic result is insufficient evidence for the whole machine. If a spare card is purchased, document the host used for its qualification and the assumptions that remain specific to the deployment chassis. Preserve the equipment identifiers with the results so a subsequent failure can be traced to the relevant assembly.

  • For an existing fleet, capture chassis model, installed accelerator identifier, firmware baseline and approved field replacement number before contacting suppliers. Keep a separate record for each server configuration.
  • For a new deployment, request an integrated demonstration of boot, accelerator enumeration, cooling behavior and representative workload operation using the proposed complete configuration.
  • For a depot repair program, define tooling, trained personnel, thermal interface handling and the test fixture required to release a repaired assembly back into service.

Pricing and Lead Time

Quotation comparisons should separate accelerator material from the integration work needed to make it usable. Request the full vendor assembly number and revision, included accessories, software or support entitlements, and any installation services. A bare module offer and a complete server offer should occupy different lines in the comparison sheet even when both contain H100 hardware.

Ask for a written description of condition and previous use where relevant. Agree the inspection evidence and functional acceptance window before purchase. For preconfigured systems, establish whether the quoted firmware and driver baseline is preserved at shipment and what happens if qualification requires an update. Do not infer such services from a product photograph or a general reference to testing.

Schedule questions should distinguish material preparation, integration, acceptance testing and transport. A seller's dispatch estimate says little about when an engineering team can release the system for workloads. Include these activities in the deployment schedule and obtain actual commercial commitments from the offer; this guide provides no live price or delivery estimate.

Alternatives and Change Control

An alternative H100 form factor represents a system change, not merely a purchasing change. Review the destination chassis, mechanical envelope, power path, cooling system, interconnect topology and software qualification. Require evidence for the exact combination proposed by the system manufacturer. Do not translate a family-level feature into a claim that an arbitrary card or module is supported.

A complete replacement server may be worth evaluating when the existing platform cannot accept the desired accelerator assembly. Compare the operational consequence: rack integration, network attachment, workload migration and maintenance coverage. The alternative with the lowest individual accelerator quotation can still demand more engineering work before it becomes an accepted production asset.

Keep semiconductor repair purchases on a separate path. A clock buffer, regulator or memory component is not an H100 service spare simply because it belongs to a category used in servers. Establish its actual board position and approved bill of materials before sourcing it. This avoids confusing a component engineering project with the acquisition of an accelerator card that already carries a system-level qualification history.

FAQ

Can I replace an H100 SXM module with an H100 PCIe card?

Do not assume that change is supported. They represent different physical implementations. Ask the server manufacturer for an approved configuration and treat a form-factor change as a platform redesign or migration.

Does a passive accelerator card include its own complete cooling system?

A passive heat sink depends on the host cooling arrangement. For H100 NVL, NVIDIA explicitly requires system airflow. Obtain the chassis-specific thermal requirements and validate them in the assembled server.

What identifier belongs on a spare-parts purchase order?

Use the approved field replacement or assembly number, revision and included hardware. Retain the GPU family description as context, but do not let it replace the identifier used by the service organization.

Discuss your requirements with SENICO and request a quotation

Sources