Interview Prep — Head of Software Engineering [GPU]
This is the interview surface for the role, drawn from the 12 phases. The role is senior and broad, so interviews span deep technical, systems design, and leadership — often in the same loop.
The five question banks
| File | Covers | Phases |
|---|---|---|
| 01 — GPU & Systems Deep-Dive | architecture, CUDA, compilers, drivers, runtime | 01–05 |
| 02 — Serving, Scheduling & Distributed | KV-cache, batching, GPU sharing, collectives | 06–08 |
| 03 — Platform, Portability & Security | HAL/ABI, observability, licensing, sovereign | 09–11 |
| 04 — System Design | the big design walkthroughs | all |
| 05 — Leadership & Behavioral | strategy, org, OEM, communication, STAR | 12 |
How to use this
- Self-test first: answer cold, then check against the model answer + the WARMUP it derives from. Gaps point you back to a phase.
- The bar is "derive, don't recite": every technical answer should reason from a principle (roofline, the 2V all-reduce bound, sign-vs-encrypt), not a memorized fact. The WARMUPs' Interview Q&A sections are the depth target.
- Leadership answers need a decision + its consequence + the honest limit — the Phase 12 communication discipline.
The 20 questions most likely to decide the loop
- Derive H100's ridge point and what it means for LLM decode. (P01)
- Walk me through what happens at the OS level on
cudaLaunchKernel. (P02/P04) - What does
nvccdo? PTX vs SASS, and the cold-start incident. (P03) - Why does PyTorch ship its own allocator? Reserved vs allocated. (P05)
- Compare MPS, MIG, time-slicing; deliver a "0.5 GPU" tier honestly. (P06)
- Explain PagedAttention — and why it's really an allocator. (P07/P05)
- Static vs continuous batching; what is goodput and why does it beat throughput? (P07)
- Derive ring all-reduce's cost; why is it bandwidth-optimal and N-independent? (P08)
- Why must tensor parallelism stay on NVLink? With numbers. (P08)
- Design a hardware-abstraction layer; walk the API/ABI boundary. (P09)
- How do you avoid a lowest-common-denominator abstraction? (P09)
- How do you version a plugin ABI so a year-old vendor backend still loads? (P09)
- What GPU metrics do you alert on, and why is "GPU utilization" misleading? (P10)
- Design a safe release process for a GPU serving platform. (P10)
- Design a license system for air-gapped deployment; sign vs encrypt? (P11)
- Anti-tamper self-checks — what do they buy, how are they defeated? (P11)
- Build vs adopt: your own compiler layer or Triton/MLIR? Defend it. (P03/P09)
- How would you execute "decouple software from hardware"? (P09/P12)
- How do you structure and hire a GPU platform team? (P12)
- Sell me the sovereign-AI value proposition. (P12)
Each maps to a WARMUP's Interview Q&A with a full principal-level answer.
Format expectations for this role
- Technical screens: whiteboard a kernel/scheduler/allocator; reason about a roofline; debug a "slow training" or "slow serving" scenario.
- System design: the system-design walkthroughs (GPU cloud, serving platform, HAL, licensing-at-scale).
- Leadership: strategy memos, org design, OEM/partner scenarios, and STAR behavioral stories — see 05.
- Hands-on credibility check: expect to be asked to actually reason about code you'd review — the labs are your evidence you can.