GPU Design · All levels

GPU Physical Design & Power: Tricky Q&A

Senior interview and review questions for GPU Physical Design & Power.

Section Q&A bank

Use these drills after completing all topics in GPU Physical Design & Power. Answer with workload context, mechanism proof, artifact, owner, and release decision.

Why is SM cluster placement tightly coupled to memory-system performance?

diagram
[INT][GPU][GPU-PHYSICAL-DESIGN]

Q: Why is SM cluster placement tightly coupled to memory-system performance?

A:
Floorplan distance and routing congestion shape latency and bandwidth consistency between SM arrays, caches, and memory interfaces.

FOLLOW-UP TRAP: Optimizing compute floorplan independent of memory proximity.

How do thermal hotspots distort timing closure assumptions?

diagram
[INT][GPU][GPU-PHYSICAL-DESIGN]

Q: How do thermal hotspots distort timing closure assumptions?

A:
Temperature gradients shift delay and skew behavior, causing corner-specific path failures not visible in uniform thermal assumptions.

FOLLOW-UP TRAP: Relying on nominal thermal models for final closure.

What confirms DVFS policy is helping instead of masking design issues?

diagram
[INT][GPU][GPU-PHYSICAL-DESIGN]

Q: What confirms DVFS policy is helping instead of masking design issues?

A:
Performance-per-watt improves while throttle duty cycle and reliability margins remain within signed guardbands across stress workloads.

FOLLOW-UP TRAP: Judging DVFS by average frequency alone.

When should clock/power partitioning be revisited?

diagram
[INT][GPU][GPU-PHYSICAL-DESIGN]

Q: When should clock/power partitioning be revisited?

A:
When crossing violations, control complexity, or wake/transition instability repeatedly block timing or integration closure.

FOLLOW-UP TRAP: Adding constraints without reevaluating partition boundaries.

Q&A drill guide

diagram
WORKLOAD -> BOTTLENECK -> METRIC -> CHANGE -> REGRESSION

Sketch while answering

diagram
GPU PD VIEW

HBM edges + SM clusters + cache rings + power/clock grid

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.