Hardware interview practice

Localize an intermittent HBM lane failure

HardPost-Silicon ValidationPYTHON

HBM traffic shows rare CRC failures that may correlate with a physical lane, address region, traffic pattern, or temperature. Implement the complete balanced experiment, numerical failure threshold, first-run classification precedence, and separate power-cycle repeat report.

Starting point

Question code

lanes 0..7
regions A,B,C,D
patterns 0x00,0xFF,0xAA,0x55
temperatures 25 C,100 C
run 10^6 transfers per tuple

Reviewed example

Work through one case

Input

Lane 3 has at least 10 errors in all four regions and all four patterns at both temperatures; every other tuple passes

Expected output

classification=LANE; all 32 lane-3 tuples appear in the repeat plan

Lane 3 contains every required 3-region by 3-pattern Cartesian subset at both temperatures, so the highest-precedence LANE rule matches.

What to cover

Requirements

Execute every lane-region-pattern-temperature tuple with equal transfer count and retain raw error records plus total transfers.
Mark a tuple failing when it has at least 10 errors per 10^6 transfers.
Apply this precedence with complete Cartesian subsets: LANE means one lane across any three regions, any three patterns, and both temperatures; REGION means one region across any six lanes, any three patterns, and both temperatures; PATTERN means one pattern across any six lanes, any three regions, and both temperatures; THERMAL means no failures at 25 C plus a complete 100 C subset spanning any four lanes, any two regions, and any two patterns; otherwise return MIXED.
Repeat every first-run failing tuple after a power cycle and report repeat results separately without changing the first-run classification.