Computer Architecture · All levels

Architecture Interview Q&A Bank

25 cross-topic senior architecture questions with traps.

Cross-topic Q&A

Answer with metric, mechanism, experiment, and validation.

Workload IPC dropped 15% after cache change — first three checks?

diagram
[INT][ARCH][BANK]

Q: Workload IPC dropped 15% after cache change — first three checks?

A:
MPKI by level, L2 bandwidth, stall breakdown (frontend/backend/memory).

FOLLOW-UP TRAP: Assume larger cache always helps.

NoC deadlock in simulation — what do you ask first?

diagram
[INT][ARCH][BANK]

Q: NoC deadlock in simulation — what do you ask first?

A:
Topology, routing algorithm, VC count, credit loop, QoS starvation.

FOLLOW-UP TRAP: Add bandwidth without root cause.

MESI vs MOESI tradeoff?

diagram
[INT][ARCH][BANK]

Q: MESI vs MOESI tradeoff?

A:
MOESI allows shared dirty in cache — reduces off-chip traffic at coherency complexity cost.

FOLLOW-UP TRAP: Name states without traffic story.

When add another cache level?

diagram
[INT][ARCH][BANK]

Q: When add another cache level?

A:
When MPKI × miss penalty × frequency justifies area/power and access latency budget.

FOLLOW-UP TRAP: More levels always better.

Branch predictor wrong — symptom?

diagram
[INT][ARCH][BANK]

Q: Branch predictor wrong — symptom?

A:
Frontend bubbles, fetch/decode stall increase, IPC drop on branch-heavy workload.

FOLLOW-UP TRAP: Only mention mispredict rate without pipeline impact.

Accelerator starved — debug order?

diagram
[INT][ARCH][BANK]

Q: Accelerator starved — debug order?

A:
DMA bandwidth, memory contention, kernel launch overhead, coherency invalidation rate.

FOLLOW-UP TRAP: Clock the accelerator faster first.

Roofline says memory-bound — fix?

diagram
[INT][ARCH][BANK]

Q: Roofline says memory-bound — fix?

A:
Improve locality, prefetch, bandwidth, or reduce arithmetic intensity requirement.

FOLLOW-UP TRAP: Wider issue width.

ACE vs CHI at high level?

diagram
[INT][ARCH][BANK]

Q: ACE vs CHI at high level?

A:
Both support coherent interconnect; CHI is newer generation with different channel/message model — know your SoC IP.

FOLLOW-UP TRAP: Treat as interchangeable buzzwords.

Out-of-order vs in-order for embedded?

diagram
[INT][ARCH][BANK]

Q: Out-of-order vs in-order for embedded?

A:
Trade IPC vs power/area/verification — cite workload and real-time determinism needs.

FOLLOW-UP TRAP: OOO always wins.

Memory ordering bug symptom?

diagram
[INT][ARCH][BANK]

Q: Memory ordering bug symptom?

A:
Rare SW failures, litmus test failures, only on weak ordering cores.

FOLLOW-UP TRAP: Debug as timing issue.

PMU shows L3 miss storm — action?

diagram
[INT][ARCH][BANK]

Q: PMU shows L3 miss storm — action?

A:
Profile working set, page allocation, sharing — consider partition or cache policy.

FOLLOW-UP TRAP: Blindly increase L3 size.

Floorplan hurts memory latency — arch response?

diagram
[INT][ARCH][BANK]

Q: Floorplan hurts memory latency — arch response?

A:
Negotiate macro placement, NoC topology, or accept frequency target change with data.

FOLLOW-UP TRAP: Ignore PD feedback.

SMT worth it?

diagram
[INT][ARCH][BANK]

Q: SMT worth it?

A:
When frontend/backend underutilized and thread mix benefits — costs area, verification, security.

FOLLOW-UP TRAP: Double threads = double IPC.

Coherency traffic dominates — fixes?

diagram
[INT][ARCH][BANK]

Q: Coherency traffic dominates — fixes?

A:
False sharing reduction, directory vs snoop fit, cache line size, data layout.

FOLLOW-UP TRAP: Faster NoC only.

Explain TLB miss impact?

diagram
[INT][ARCH][BANK]

Q: Explain TLB miss impact?

A:
Page walk latency, pipeline stall, may dominate in sparse memory workloads.

FOLLOW-UP TRAP: Ignore MMU in perf analysis.

Staff architect vs senior?

diagram
[INT][ARCH][BANK]

Q: Staff architect vs senior?

A:
Staff owns cross-block budgets, methodology, signoff narrative, and escalation.

FOLLOW-UP TRAP: More technical trivia.

NPU integration top risk?

diagram
[INT][ARCH][BANK]

Q: NPU integration top risk?

A:
Memory bandwidth and coherency with CPU — define ownership and bandwidth contract early.

FOLLOW-UP TRAP: Peak TOPs only.

When pipeline deeper?

diagram
[INT][ARCH][BANK]

Q: When pipeline deeper?

A:
When target frequency and branch predictability support — watch branch and cache penalties.

FOLLOW-UP TRAP: Depth without frequency goal.

Chip PPA negotiation example?

diagram
[INT][ARCH][BANK]

Q: Chip PPA negotiation example?

A:
Present metric dashboard: perf/W, area mm², NoC BW — options with owners.

FOLLOW-UP TRAP: Single metric advocacy.

Architecture signoff artifact?

diagram
[INT][ARCH][BANK]

Q: Architecture signoff artifact?

A:
Workload matrix, perf/Power budgets, verification plan, PD constraints doc.

FOLLOW-UP TRAP: PowerPoint only.

Prefetcher harmful case?

diagram
[INT][ARCH][BANK]

Q: Prefetcher harmful case?

A:
Pollution on streaming or irregular access — measure MPKI and bandwidth.

FOLLOW-UP TRAP: Prefetch always helps.

QoS on NoC — why?

diagram
[INT][ARCH][BANK]

Q: QoS on NoC — why?

A:
Prevent latency-critical traffic starvation by best-effort bulk DMA.

FOLLOW-UP TRAP: Fair arbitration is enough.

Cache inclusion policy debate?

diagram
[INT][ARCH][BANK]

Q: Cache inclusion policy debate?

A:
Inclusion simplifies coherence but can cause eviction traffic — workload dependent.

FOLLOW-UP TRAP: Always inclusive.

Explain memory bandwidth roof?

diagram
[INT][ARCH][BANK]

Q: Explain memory bandwidth roof?

A:
Pin count, DRAM type, NoC links, coherency overhead — numeric estimate.

FOLLOW-UP TRAP: Peak datasheet bandwidth only.

Post-silicon perf mismatch?

diagram
[INT][ARCH][BANK]

Q: Post-silicon perf mismatch?

A:
Workload diff, thermal throttle, binning, PMU setup, trace alignment.

FOLLOW-UP TRAP: Blame RTL immediately.

Q&A drill guide

diagram
MECHANISM  METRIC  EXPERIMENT  TRADEOFF  VALIDATION

Sketch while answering

diagram
WHITEBOARD ANSWER SHAPE

1. Workload and metric
2. Block diagram
3. Bottleneck hypothesis tree
4. Experiment plan
5. Design options and tradeoffs
6. Validation / regression

If you skip step 1, the rest sounds generic.

Key takeaways

  • Connect every architecture claim to a workload and measurable metric.

  • State verification and PPA impact before proposing design changes.

Common pitfalls

  • Feature-driven design without MPKI/IPC/bandwidth evidence.

  • Ignoring coherency and NoC traffic in cache and accelerator sizing.