AI server certification before dispatch

GPU qualification and server certification

Qualification of an individual GPU card and certification of the complete server are separate processes. They use different tools, have different cooling conditions and assess different subjects. A card’s qualification result on a test bench does not replace verification of its operation in a complete 2U server configuration beneath an air shroud, in an airflow channel with specific fans.

We therefore certify the assembled machine separately, and measurement results from the two processes should not be compared directly. How we qualify individual cards: how we check every used GPU accelerator.

Conditions for starting a certification run

1. Inventory with serial numbers collected before testing — processors, memory modules by slot, drives by bay, GPU cards, controllers. Without serial numbers, the result cannot be unambiguously assigned to a specific unit, and firmware servicing operations irreversibly overwrite the inventory.
2. Recorded BIOS power profile— the profile can change performance results by a factor of several. Without recording it, the figures in the certificate would not be reproducible.
3. PCIe link width and speed of each card confirmed from LnkSta — the management interface may report full width when the link is operating at a reduced width. A reduced width detected after more than ten hours of testing requires the entire run to be repeated.
4. Complete GPU cooling equipment — air shrouds, risers and power supplies of the appropriate wattage. With an incomplete airflow channel, passive GPU cards may lower their clock speeds to shed heat (thermal throttling), which invalidates the result.
5. Thermal safety thresholds — GPU thresholds are determined from parameters reported by the card; CPU temperature is monitored against a defined safety threshold.
6. Hardware event log baseline recorded before load testing — events that have occurred since the start of the run are analysed after testing.

Conformity with the shipping configuration

Certification checks more than whether the server works. The system automatically compares the hardware actually detected with the profile assigned to the specific unit intended for dispatch. Checks include:

• processors
• memory
• GPU cards
• drives and storage controller
• power supplies
• network interfaces

A mismatch with the required profile prevents a positive result.

Power supplies and network interfaces

Before load testing begins, the system automatically checks the presence, condition and redundancy of power supplies and whether their number and wattage match the server’s profile. The required network interfaces, port count and speed class are also verified. A missing required component or a critical state prevents a CERTIFICATION result.

PCIe link of each GPU card

• Maximum interface capabilities (capability) and negotiated link parameters are separate values.We read both: the parameters supported by the card’s interface and those actually negotiated with the slot (LnkSta), because the management controller may report full width when the link is operating at a reduced width.
• A card with x16 capability that negotiates only x8 has a degraded link, rather than a normal variation. Such a run cannot receive a CERTIFICATION result.
• We map cards by unit identity — serial number and PCI address — rather than solely by their order in nvidia-smi, which may not correspond to slot order. Each measurement result is assigned to a specific card in a specific slot.

Certification does not modify drive data

We start the run from live media, with the system running in RAM. Certification does not install an operating system, partition or format drives, or perform tests that write user data or test patterns to the data area of the target drives. We assess drive condition using SMART and the self-test available for the drive type, firmware and RAID controller. The drive itself performs the self-test, and its firmware may update its own SMART diagnostic log in the process. If a full long self-test is not supported or its result cannot be obtained reliably through the controller layer, the procedure takes this into account in the final classification of the run. The hardware configuration during testing is the same configuration intended for delivery — any change to the installed memory modules, drives, cards or risers invalidates the run and requires it to be repeated.

Certification run

The sequence is mandatory. An interrupted or incomplete run cannot form the basis of a certificate — to obtain certification, we perform a new, complete run from the beginning. We do not accept a partial result as evidence of correct operation.

Monitoring during the run

1. Thermal watchdog. CPU and GPU loads are supervised by a thermal watchdog that can automatically stop the test if a safety threshold is exceeded.
2. Sensor sampler records temperatures, power consumption, clock speeds and the reasons the clocks were lowered (throttle reasons) every 10 s throughout the run.
3. Throttling due to temperature is treated as a condition that disqualifies the run. Throttling due to power limits is expected and is not grounds for rejection — the report distinguishes between the two cases.

Not every completed test results in certification

The final assessment is divided into three independent layers:

• Hardware Result — hardware test results.
• Evidence completeness (Evidence Pipeline) — completeness and correctness of the evidence.
• Certification Result — final classification of the run as certification, validation or failure.

The procedure can conclude a run with one of three results:

• CERTIFICATION — certification: the hardware, configuration and evidence meet all requirements.
• VALIDATION — validation: the hardware may operate correctly, but at least one formal or evidence-related condition prevents a certificate from being issued.
• FAIL — failure: a hardware problem, configuration mismatch or other disqualifying condition has been detected.

The document generator does not allow a VALIDATION or FAIL result to be changed manually to CERTIFICATION. A certificate is issued only for a run that ends with a CERTIFICATION result.

A server with a VALIDATION result is not dispatched as a certified unit. The validation report is an internal document and does not replace the customer certificate — once the cause has been resolved, we perform a new, complete run that must end with a CERTIFICATION result.

Resulting documentation

• Every value presented in the documentation must be supported by retained evidence from the run. Values are not entered manually from console readings or operator notes.
• UTC time only, without local time — ensuring that timestamps can be compared between runs.
• Archive checksum verified after copying to the repository, as well as after creation.
• The certificate does not state causes that have not been confirmed by measurement.
• Technical terminology is retained in its original form, without descriptive translations, so that it can be verified.

What certification does not cover

We sell used hardware from the secondary market. The certificate confirms the machine’s condition at the time of the run, and the retained archive enables subsequent verification of the results. The certificate is not a manufacturer’s warranty. The seller’s warranty applies on the terms stated at the time of sale.

Performance figures in the certificate are not thresholds. Performance stages report the value measured on this machine, with this BIOS profile and these tool parameters. Two benchmarks with the same name, but different versions and parameters, can produce significantly different results on the same processor — comparing our result with that of another tool therefore requires both configurations to be stated.

The self-test supplied with the server is a different tool from certification. It confirms after delivery that components are present and operational; it does not assess performance or issue a certificate. Results from the acceptance self-test and the certification run should not be compared directly, as they use different tools and have different scopes.