Read first: Read the mechanism and worked trace first
A reproducible memory failure report must distinguish workload, configuration and observation. “Memory test failed after boot” cannot reveal whether two machines ran the same work.
Preserve enough context for another engineer to replay the failure — decision record
Apply the stated local assumptions; record actual platform values before using this as an implementation checklist.
Decision / observation | Required evidence | Consequence |
|---|---|---|
Configuration | Part/mode, active timings, clock, mapping and training result | A nominal data rate alone is insufficient |
Workload | Initial data, access sequence, address range and concurrency | Preserve failing byte/address and sequence identity |
Conditions | Temperature/voltage observations and reset history | Avoid comparing different corners as if only firmware changed |
Result | First mismatch, command/lane evidence and performed work | A passing shorter run is not an equivalent replay |
The configuration, workload, conditions, performed work and checked result, compared with the failing replay. Otherwise the result may not exercise the original failure.
Memory debug involves silicon (controller/PHY), firmware (training, MRW), board (SI, routing), and software (address map, cache policy). The review memo must name which layer owns the evidence.
SI team: eye diagrams, channel loss, crosstalk.
Firmware: training sequence, mode registers, QoS settings.
Software: memory attributes, DMA buffer alignment, interleave.