GPU Design · All levels

Graphics Pipeline Architecture: Tricky Q&A

Senior interview and review questions for Graphics Pipeline Architecture.

Section Q&A bank

Use these drills after completing all topics in Graphics Pipeline Architecture. Answer with workload context, mechanism proof, artifact, owner, and release decision.

Why can frame-time spikes appear after async compute enablement?

diagram
[INT][GPU][GRAPHICS-PIPELINE]

Q: Why can frame-time spikes appear after async compute enablement?

A:
Queue arbitration and shared back-end resources can create contention between fragment and compute workloads under thermal and bandwidth pressure.

FOLLOW-UP TRAP: Treating graphics and compute queues as independent systems.

How do you verify early-Z is delivering real savings?

diagram
[INT][GPU][GRAPHICS-PIPELINE]

Q: How do you verify early-Z is delivering real savings?

A:
Correlate kill-rate improvement with fragment workload reduction and lower back-end memory traffic on the same scene set.

FOLLOW-UP TRAP: Accepting depth-test counters without downstream impact checks.

What differentiates a raster bottleneck from an ROP bottleneck?

diagram
[INT][GPU][GRAPHICS-PIPELINE]

Q: What differentiates a raster bottleneck from an ROP bottleneck?

A:
Raster limits show primitive/fragment setup pressure, while ROP limits show blend queue buildup and color/depth write saturation.

FOLLOW-UP TRAP: Grouping all back-end slowdowns under generic fill-rate issues.

When is tile-based deferred rendering likely to help most?

diagram
[INT][GPU][GRAPHICS-PIPELINE]

Q: When is tile-based deferred rendering likely to help most?

A:
Scenes with high overdraw and locality benefit most from on-chip tile reuse before external memory resolve.

FOLLOW-UP TRAP: Assuming TBDR helps equally across all content mixes.

Q&A drill guide

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

Sketch while answering

diagram
GRAPHICS PIPELINE

vertex -> tessellation -> raster -> fragment -> ROP/blend

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.