GPU Design · All levels
Scenario: Occupancy Mystery
A team increases block size to raise occupancy, but real throughput drops. Registers per thread and shared-memory pressure changed after a compiler update.
Scenario
A team increases block size to raise occupancy, but real throughput drops. Registers per thread and shared-memory pressure changed after a compiler update.
OBSERVED METRIC
kernel latency regresses despite higher theoretical occupancy
45-MINUTE INTERVIEW FLOW
0-5: scope workload and KPI
5-15: map SIMT/memory/scheduler mechanism
15-25: isolate first failing evidence
25-35: propose bounded fix with owner
35-45: state validation matrix + rollbackCommon pitfalls
Assume occupancy alone predicts performance
Ignore register spilling and memory replay counters
Skip instruction mix and dependency stall analysis
Scenario debrief
Score candidate response on workload framing, bottleneck proof, mitigation choice, and regression discipline.
kernel intent -> hardware execution -> measured outcomethroughput trendDebrief prompts
Which resource bottleneck appears first in evidence?
What hardware or software lever changes the bottleneck?
What reproducible benchmark gate proves closure?
Key takeaways
Always pair micro-kernel metrics with end-to-end workload impact.
Lock toolchain, driver, and launch metadata before comparing performance results.
Common pitfalls
Optimizing occupancy without checking memory-system saturation.
Comparing profiler captures from different driver or compiler builds.
Declaring wins without reproducible accuracy and performance gates.