County health sees one PSU drop voltage only when AI training peaks, while sibling supplies stay healthy. What conclusion follows?
Select an answer to reveal the explanation.
Short Explanation
When you see one PSU wilt under an AI training peak while its siblings stay strong, you’ve found a load-sensitive outlier. Tag that PSU as the faulty power-supply FRU.
Full Explanation
When county health observes one PSU drop voltage only at AI training peaks while sibling supplies stay healthy, the load-sensitive PSU is the faulty power-supply FRU. Differential behavior under the same training load isolates the failing supply without blaming unrelated subsystems; a unit that holds at idle but sags under peak GPU draw is showing classic load-sensitive FRU behavior. Filesystem mount settings cannot explain one PSU sagging while sibling supplies hold voltage, because storage client configuration is a different fault domain. Concluding that all PSUs are defective merely because load rose ignores the healthy siblings under that same load and could waste redundant-supply spares. GPU MIG profiles change compute partitioning but do not account for one supply’s voltage collapse. Exam caveat: verify the sagging rail with PDU or server power telemetry during a controlled training peak, ruling out a shared facility feed or PDU phase issue before an RMA. Practical check: trend per-PSU voltage and input health during a representative peak, identify the outlier, replace that FRU according to redundant-PSU guidance, then repeat the peak and verify both supplies hold voltage.