One question on this site is reported in NVIDIA software engineer interviews. Each is a guided run through requirements, back-of-envelope numbers, an interactive design, user flows and deep dives with interviewer follow-ups. Patterns covered: gpu-batching, streaming, scheduling, kv-cache.
Run an AI mock NVIDIA interview →One of these questions, run as a live round in that style, with a debrief at the end.
Attribution comes from public interview guides and candidate reports, not from NVIDIA. It indicates where a question has been asked, nothing more.
Serve a large language model behind an API: queue the requests, batch them onto scarce GPUs, stream tokens back, and keep every accelerator busy without blowing the latency budget.
45–60 minutes, often close to accelerators: serving, scheduling or data movement for large models.
Quoting hardware names without the arithmetic. The numbers are the answer here.
From public interview guides and candidate reports, not from NVIDIA. Formats change; treat it as a prior, not a promise.
Large-company loops run 45 to 60 minutes, expect numbers, and spend most of the time on deep dives. Work each question above in practice mode until you can produce the requirements, the estimates and the flows without looking, then read the interview framework guide for how to structure the hour. The other questions on the site teach patterns that show up in NVIDIA interviews under different names.