Design a GPU Cluster Job Scheduler
A job scheduler takes work from many users and decides where and when each piece runs. Here the machines are GPU hosts, usually 8 GPUs to a box, and the work is containers: training runs, evaluations, batch inference, test suites. The user says what the job needs and the scheduler finds a place for it, starts it, watches it and cleans up after it.
Versions of this prompt show up in NVIDIA technical screens and onsites, for SWE and Senior SWE roles on infrastructure and cloud teams. In an October 2026 onsite it came up as a task service that runs each task when it's due. A senior first round in November 2026 asked for a job scheduler, with questions on execution guarantees and metrics logging. A mid-2026 screen framed it as a control plane that tracks live host state across a large fleet and dispatches jobs. An older version, from a 2023 onsite, asked for an interface for managing and running jobs on various GPUs. Under every version sit two hard problems. Each job has to run once even when hosts, networks and schedulers fail, and the fleet's state changes thousands of times a second. The single-machine cousin, one queue feeding a worker thread per GPU, is worked in the low-level design lesson. This one spreads the same idea across a fleet.
NVIDIAPremium lesson · System design round
The NVIDIA design, worked end to end.
- The API and data modelSystem design · Frequently asked
- The design, drawn out2 diagrams
- 5 deep divesWith the tradeoffs
$49 a year · all 107 companies · instant access
139 NVIDIA interview reports · guide updated
NVIDIA interview guideKnow what NVIDIA asks before you walk in.
19 coding · 5 system design · 25 behavioral questions
Get their questions first.
InterviewLegend is independent and not affiliated with or endorsed by any company it covers.