SENICO Electronics

Component sourcing and inventory guides

MI300X HBM Acceptance: ECC History and Retired Memory Pages

Build an MI300X memory acceptance record that distinguishes ECC events, retired pages, tool support and repeatable workload observations.

Quick Answer

MI300X HBM acceptance should include memory health evidence as well as capacity. AMD lists ECC, page retirement and page avoidance among the accelerator's reliability features. These mechanisms describe how the platform handles memory conditions; they do not establish that a particular offered unit has an acceptable service history. Agree the condition criteria, capture a baseline and compare it with observations after qualification.

Treat unknown telemetry as unknown. An unsupported query, a missing software privilege or an empty report should not be rewritten as zero errors. The acceptance record must say what was measured, which tool produced it and whether the requested data was successfully retrieved.

Key Specifications and Decision Points

AMD's management documentation provides interfaces for ECC and retired-memory information, with details varying by software release and environment. Its current AMD SMI material distinguishes retired or reserved pages from general device information. Select the documentation matching the installed tool before deciding which fields belong in an inspection report.

Capacity, corrected events and retired pages are related observations, not interchangeable measurements. Preserve their labels and units as returned by the relevant interface. If a seller supplies a screenshot without tool version or device identity, ask for a more complete record before using it to approve the equipment.

  • Identify the accelerator using a stable hardware identifier and record the host location. Device numbering can change, so a bare index is inadequate for long-term comparison.
  • Record driver, firmware, operating system and management-tool versions. Include the exact query outcome, including unsupported or unavailable fields, without translating failures into healthy values.
  • Separate cumulative history from changes observed during the test. A nonzero baseline and a newly increasing count pose different engineering questions.

Application and Acceptance Review

Start with the system idle and save the available memory health report. Run the agreed workload under controlled cooling and power conditions, then capture the same report again. Compare changes for the same physical accelerator. This approach does not prescribe a universal rejection threshold; it produces evidence that the buyer and system supplier can evaluate against their agreed policy.

Do not clear logs or alter reliability settings merely to produce a cleaner acceptance screen. If a documented service procedure requires a reset, save the original evidence and record the reason, operator and resulting state. The distinction between a fresh baseline and erased history matters when equipment changes ownership or returns from repair.

For a multi-accelerator server, collect records for every physical device. One passing module does not establish the health of the other modules sharing the baseboard, cooling environment or workload.

  • Include workload initialization and sustained operation so the observation covers more than a brief enumeration. State the duration and the memory usage actually exercised.
  • Log errors and operating conditions together. A thermal or power incident during testing may be relevant to diagnosis even when the eventual symptom appears in memory telemetry.
  • Repeat a questionable result using the supported diagnostic procedure. Preserve the first result as evidence rather than deleting it when a later attempt looks better.

Pricing and Lead Time

Place memory-condition expectations in the RFQ before asking sellers for diagnostic evidence. The commercial requirement should define the type of equipment, relevant prior use disclosure, test procedure and handling of discrepancies. Avoid unsupported claims that a generic test label proves an unlimited operating lifetime or that a seller's acceptance criteria automatically match yours.

Ask which party performs the test and whether results will identify the exact shipped units. Where a report predates shipment, request confirmation that the equipment configuration and identity remain the same. If the buyer will run acceptance after arrival, define the reporting window and escalation route for an unexpected result.

Pricing should be compared after the condition and evidence scope is understood. A report that omits retired-page information is not necessarily evidence of a fault, but it is also not equivalent to a complete health record. Obtain current commercial terms for the agreed condition and inspection scope rather than assuming all MI300X offers are alike.

Alternatives and Change Control

If a query is unsupported, the next step is to establish why. Check the supported driver and management stack, execution environment and documentation for that release. A bare-metal diagnostic path may expose information unavailable through a virtualized interface. Ask the system supplier for its supported inspection procedure; do not infer a hardware failure solely from a tooling limitation.

When evidence points to a real memory issue, use the server vendor's service process and applicable acceptance policy. The appropriate remedy may involve an accelerator assembly rather than an individual memory component. HBM is part of the accelerator's integrated packaging and should not be presented in an RFQ as an ordinary field-upgradable DIMM.

Changing accelerator families does not remove the need for a health baseline. Define comparable evidence categories for the alternative platform while retaining vendor-specific terminology and thresholds. The objective is a traceable decision about the delivered hardware, not a cosmetic report in which every vendor's diagnostics have been forced into identical fields with lost meaning.

FAQ

Does a corrected ECC event automatically mean an accelerator must be rejected?

Use the manufacturer and system supplier guidance together with the agreed acceptance policy. Record the event type, history and changes during testing. This guide does not invent a universal rejection count.

Can an unsupported memory-health query be reported as zero?

No. Record that the field was unavailable and investigate the supported diagnostic route. A measurement failure and a successfully measured zero are different evidence states.

Can a retired HBM page be addressed by ordering a DIMM?

HBM on MI300X is part of the accelerator assembly architecture. Follow the approved service procedure for that equipment. Do not treat accelerator HBM as a replaceable server memory module.

Discuss your requirements with SENICO and request a quotation

Sources