SENICO Electronics

Component sourcing and inventory guides

MI300X OAM and UBB: Accepting an Eight-GPU Assembly

Separate MI300X OAM modules, eight-GPU UBB assemblies and complete servers when defining serial records, connectivity checks and system acceptance.

Quick Answer

Accept an MI300X installation at the assembly level actually purchased. AMD describes MI300X as an OAM accelerator and its eight-accelerator platform as a UBB assembly. A complete AI server adds host processors, networking, storage, power distribution, cooling and system firmware. These scopes should not be merged into a single vague line reading MI300X hardware.

The buyer should receive a configuration record connecting each accelerator to its baseboard and server. That record makes incoming inspection, functional acceptance and later maintenance meaningful. An OAM module is an accelerator assembly, not a socketed desktop graphics card or a discrete component for ordinary board assembly.

Key Specifications and Decision Points

AMD specifies 192 GB of HBM3 per MI300X accelerator and describes a UBB platform containing eight modules, with 1.5 TB aggregate memory. Its platform material also identifies PCIe host connections and GPU interconnects. Use those architectural facts to prepare the inspection map, while obtaining installation details from the supplier of the actual server.

An eight-device headline does not settle how the host routes traffic to each accelerator. Request the platform topology diagram and use it as an expected result during qualification. A connectivity error should be traceable to a named path rather than a generic observation that one GPU seems slow.

  • Give every module a recorded position, serial identity and associated baseboard identity. Preserve a photograph or diagram of the position naming convention used by the service team.
  • Define what ships with the UBB: cooling assemblies, retention hardware, cables, power interfaces and any required management connections. Resolve omissions before material is dispatched.
  • Separate physical device count from the devices presented by a software partitioning configuration. Record the selected operating mode alongside enumeration evidence.

Application and Acceptance Review

Build acceptance in stages so a fault can be localized. Begin with a visual and identity inspection, continue with firmware and driver enumeration, then evaluate interconnect behavior and representative workload operation. Preserve the result of each stage instead of reducing the whole process to one pass indicator. This makes it possible to distinguish a packaging discrepancy from a system configuration defect.

For a replacement baseboard, also exercise service recovery: install the approved firmware baseline, restore configuration and confirm that monitoring associates each physical position with the right software device. Labels in an old monitoring dashboard can remain wrong even when every accelerator is functioning. Include a deliberate check of that mapping in the handover.

A short demonstration establishes only the tested conditions. Agree the workload duration, test environment and data retained so neither party mistakes a boot demonstration for a complete integration qualification.

  • Compare all detected physical accelerators with the receiving record. Investigate missing or unexpected identities before running a workload that might obscure the mapping.
  • Check the expected communication paths between devices and between the accelerators and host. Include the test tool version, operating mode and relevant topology output.
  • Observe temperatures, power behavior and errors while the complete assembly runs an agreed workload. Use the server manufacturer's limits and procedure when deciding whether operation is acceptable.

Pricing and Lead Time

Request quotations against an explicit bill of delivered assemblies. An offer for eight loose OAM modules cannot be compared directly with an integrated UBB or a complete server. State whether the requirement is expansion, a maintenance spare or a new system build, because those uses create different dependencies and acceptance responsibilities.

Ask the supplier to identify the exact platform revision and the evidence connecting the offered configuration to its supported chassis. Include the required accessory list and the agreed reporting format. For serialized equipment, decide how identities will be recorded before delivery and verified at receiving. If substitutions are proposed, require a revised configuration record before commercial acceptance.

Separate integration milestones from transport milestones. The commercial response should say when the assembly can be inspected, when system testing is expected to finish and what date refers to dispatch. Current quantity, price and schedule require an actual offer. A manufacturer specification establishes architecture, not the condition or availability of a particular supplier's equipment.

Alternatives and Change Control

An alternative MI300X system from another integrator may use the same accelerator family but differ in cooling, host layout, management firmware and service organization. Compare the complete configurations using a shared workload and acceptance procedure. Preserve vendor-specific differences instead of forcing them into an overly simple GPU-count comparison.

A loose-module purchasing route may suit an organization with an established integration program, qualified tooling and appropriate support. A complete-server route may reduce the amount of integration the buyer must manage. Evaluate the responsibilities transferred by each offer, including who diagnoses failures spanning the UBB, host and cooling system.

Changing to a different accelerator generation creates another decision. Memory capacity and theoretical bandwidth alone cannot confirm that the existing software or rack design will remain suitable. Review model execution, precision requirements, physical integration and operational support together. Keep the MI300X acceptance record as a documented reference, while defining new tests for capabilities or failure modes introduced by the proposed platform.

FAQ

Is an MI300X OAM module the same thing as an eight-GPU UBB?

No. An OAM module is one accelerator assembly. AMD describes the platform as a baseboard hosting eight such modules. A quotation must identify which of those assembly levels is included.

Why record serial numbers by physical position?

Position mapping connects software diagnostics to hardware service actions. It helps maintenance teams replace the intended assembly and prevents a monitoring label from being mistaken for a reliable physical identity.

Does total HBM capacity prove a complete server is accepted?

Capacity is one check. The server also needs the expected device count, interconnect behavior, firmware configuration, cooling performance and workload operation. Set those acceptance criteria before comparing offers.

Discuss your requirements with SENICO and request a quotation

Sources